diff --git "a/dataset_infos.json" "b/dataset_infos.json" deleted file mode 100644--- "a/dataset_infos.json" +++ /dev/null @@ -1 +0,0 @@ -{"XNLI": {"description": "\nThe Cross-lingual Natural Language Inference (XNLI) corpus is a crowd-sourced collection of 5,000 test and\n2,500 dev pairs for the MultiNLI corpus. The pairs are annotated with textual entailment and translated into\n14 languages: French, Spanish, German, Greek, Bulgarian, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese,\nHindi, Swahili and Urdu. This results in 112.5k annotated pairs. Each premise can be associated with the\ncorresponding hypothesis in the 15 languages, summing up to more than 1.5M combinations. The corpus is made to\nevaluate how to perform inference in any language (including low-resources ones like Swahili or Urdu) when only\nEnglish NLI data is available at training time. One solution is cross-lingual sentence encoding, for which XNLI\nis an evaluation benchmark.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @InProceedings{conneau2018xnli,\n author = {Conneau, Alexis\n and Rinott, Ruty\n and Lample, Guillaume\n and Williams, Adina\n and Bowman, Samuel R.\n and Schwenk, Holger\n and Stoyanov, Veselin},\n title = {XNLI: Evaluating Cross-lingual Sentence Representations},\n booktitle = {Proceedings of the 2018 Conference on Empirical Methods\n in Natural Language Processing},\n year = {2018},\n publisher = {Association for Computational Linguistics},\n location = {Brussels, Belgium},\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://www.nyu.edu/projects/bowman/xnli/", "license": "", "features": {"language": {"dtype": "string", "id": null, "_type": "Value"}, "sentence1": {"dtype": "string", "id": null, "_type": "Value"}, "sentence2": {"dtype": "string", "id": null, "_type": "Value"}, "gold_label": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "XNLI", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 20359500, "num_examples": 75150, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 10049303, "num_examples": 37350, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/XNLI/XNLI-1.0.zip": {"num_bytes": 17865352, "checksum": "4ba1d5e1afdb7161f0f23c66dc787802ccfa8a25a3ddd3b165a35e50df346ab1"}}, "download_size": 17865352, "post_processing_size": null, "dataset_size": 30408803, "size_in_bytes": 48274155}, "tydiqa": {"description": "Gold passage task (GoldP): Given a passage that is guaranteed to contain the\n answer, predict the single contiguous span of characters that answers the question. This is more similar to\n existing reading comprehension datasets (as opposed to the information-seeking task outlined above).\n This task is constructed with two goals in mind: (1) more directly comparing with prior work and (2) providing\n a simplified way for researchers to use TyDi QA by providing compatibility with existing code for SQuAD 1.1,\n XQuAD, and MLQA. Toward these goals, the gold passage task differs from the primary task in several ways:\n only the gold answer passage is provided rather than the entire Wikipedia article;\n unanswerable questions have been discarded, similar to MLQA and XQuAD;\n we evaluate with the SQuAD 1.1 metrics like XQuAD; and\n Thai and Japanese are removed since the lack of whitespace breaks some tools.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{tydiqa,\n title = {TyDi QA: A Benchmark for Information-Seeking Question Answering in Typologically Diverse Languages},\n author = {Jonathan H. Clark and Eunsol Choi and Michael Collins and Dan Garrette and Tom Kwiatkowski and Vitaly Nikolaev and Jennimaria Palomaki}\n year = {2020},\n journal = {Transactions of the Association for Computational Linguistics}\n }\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/google-research-datasets/tydiqa", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tydiqa", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"train": {"name": "train", "num_bytes": 52948607, "num_examples": 49881, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 5006461, "num_examples": 5077, "dataset_name": "xtreme"}}, "download_checksums": {"https://storage.googleapis.com/tydiqa/v1.1/tydiqa-goldp-v1.1-train.json": {"num_bytes": 58004076, "checksum": "cefc8e09ff2548d9b10a678d3a6bbbe5bc036be543f92418819ea676c97be23b"}, "https://storage.googleapis.com/tydiqa/v1.1/tydiqa-goldp-v1.1-dev.json": {"num_bytes": 5617409, "checksum": "b286e0f34bc7f52259359989716f369b160565bd12ad8f3a3e311f9b0dbad1c0"}}, "download_size": 63621485, "post_processing_size": null, "dataset_size": 57955068, "size_in_bytes": 121576553}, "SQuAD": {"description": "Stanford Question Answering Dataset (SQuAD) is a reading comprehension dataset, consisting of questions posed by crowdworkers on a set of Wikipedia articles, where the answer to every question is a segment of text, or span, from the corresponding reading passage, or the question might be unanswerable.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{2016arXiv160605250R,\n author = {{Rajpurkar}, Pranav and {Zhang}, Jian and {Lopyrev},\n Konstantin and {Liang}, Percy},\n title = \"{SQuAD: 100,000+ Questions for Machine Comprehension of Text}\",\n journal = {arXiv e-prints},\n year = 2016,\n eid = {arXiv:1606.05250},\n pages = {arXiv:1606.05250},\n archivePrefix = {arXiv},\n eprint = {1606.05250},\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://rajpurkar.github.io/SQuAD-explorer/", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "SQuAD", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"train": {"name": "train", "num_bytes": 79317110, "num_examples": 87599, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 10472653, "num_examples": 10570, "dataset_name": "xtreme"}}, "download_checksums": {"https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v1.1.json": {"num_bytes": 30288272, "checksum": "3527663986b8295af4f7fcdff1ba1ff3f72d07d61a20f487cb238a6ef92fd955"}, "https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v1.1.json": {"num_bytes": 4854279, "checksum": "95aa6a52d5d6a735563366753ca50492a658031da74f301ac5238b03966972c9"}}, "download_size": 35142551, "post_processing_size": null, "dataset_size": 89789763, "size_in_bytes": 124932314}, "PAN-X.af": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.af", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 259709, "num_examples": 1000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 257204, "num_examples": 1000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 1321396, "num_examples": 5000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 1838309, "size_in_bytes": 235847193}, "PAN-X.ar": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.ar", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 1808303, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 1811983, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 3634136, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 7254422, "size_in_bytes": 241263306}, "PAN-X.bg": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.bg", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 2310314, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 2306158, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 4600773, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 9217245, "size_in_bytes": 243226129}, "PAN-X.bn": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.bn", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 159088, "num_examples": 1000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 159282, "num_examples": 1000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 1568845, "num_examples": 10000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 1887215, "size_in_bytes": 235896099}, "PAN-X.de": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.de", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 2381565, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 2377639, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 4762352, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 9521556, "size_in_bytes": 243530440}, "PAN-X.el": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.el", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 2533806, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 2547594, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 5063176, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 10144576, "size_in_bytes": 244153460}, "PAN-X.en": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.en", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 1920069, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 1916220, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 3823474, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 7659763, "size_in_bytes": 241668647}, "PAN-X.es": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.es", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 1592525, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 1602291, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 3199161, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 6393977, "size_in_bytes": 240402861}, "PAN-X.et": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.et", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 2030160, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 2021409, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 3023211, "num_examples": 15000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 7074780, "size_in_bytes": 241083664}, "PAN-X.eu": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.eu", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 2296335, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 2249835, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 2292327, "num_examples": 10000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 6838497, "size_in_bytes": 240847381}, "PAN-X.fa": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.fa", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 1782306, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 1770284, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 3529354, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 7081944, "size_in_bytes": 241090828}, "PAN-X.fi": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.fi", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 2131769, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 2130665, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 4273793, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 8536227, "size_in_bytes": 242545111}, "PAN-X.fr": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.fr", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 1664190, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 1675785, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 3335424, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 6675399, "size_in_bytes": 240684283}, "PAN-X.he": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.he", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 2332760, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 2318756, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 4667100, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 9318616, "size_in_bytes": 243327500}, "PAN-X.hi": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.hi", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 190671, "num_examples": 1000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 196190, "num_examples": 1000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 964212, "num_examples": 5000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 1351073, "size_in_bytes": 235359957}, "PAN-X.hu": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.hu", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 2211851, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 2249779, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 4499914, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 8961544, "size_in_bytes": 242970428}, "PAN-X.id": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.id", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 1537979, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 1536879, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 3084007, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 6158865, "size_in_bytes": 240167749}, "PAN-X.it": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.it", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 1908529, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 1928408, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 3874663, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 7711600, "size_in_bytes": 241720484}, "PAN-X.ja": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.ja", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 6323003, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 6448960, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 12670401, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 25442364, "size_in_bytes": 259451248}, "PAN-X.jv": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.jv", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 14600, "num_examples": 100, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 16917, "num_examples": 100, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 16106, "num_examples": 100, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 47623, "size_in_bytes": 234056507}, "PAN-X.ka": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.ka", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 2806901, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 2824641, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 2777362, "num_examples": 10000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 8408904, "size_in_bytes": 242417788}, "PAN-X.kk": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.kk", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 238109, "num_examples": 1000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 236724, "num_examples": 1000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 240276, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 715109, "size_in_bytes": 234723993}, "PAN-X.ko": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.ko", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 2138167, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 2138294, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 4284733, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 8561194, "size_in_bytes": 242570078}, "PAN-X.ml": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.ml", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 290755, "num_examples": 1000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 276926, "num_examples": 1000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 2865204, "num_examples": 10000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 3432885, "size_in_bytes": 237441769}, "PAN-X.mr": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.mr", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 245358, "num_examples": 1000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 255904, "num_examples": 1000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 1248259, "num_examples": 5000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 1749521, "size_in_bytes": 235758405}, "PAN-X.ms": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.ms", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 147515, "num_examples": 1000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 147168, "num_examples": 1000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 2965048, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 3259731, "size_in_bytes": 237268615}, "PAN-X.my": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.my", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 40428, "num_examples": 100, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 37366, "num_examples": 100, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 32735, "num_examples": 100, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 110529, "size_in_bytes": 234119413}, "PAN-X.nl": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.nl", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 2016856, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 2038638, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 4062189, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 8117683, "size_in_bytes": 242126567}, "PAN-X.pt": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.pt", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 1575141, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 1562625, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 3149283, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 6287049, "size_in_bytes": 240295933}, "PAN-X.ru": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.ru", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 2053169, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 2074145, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 4121791, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 8249105, "size_in_bytes": 242257989}, "PAN-X.sw": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.sw", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 136368, "num_examples": 1000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 140231, "num_examples": 1000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 135911, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 412510, "size_in_bytes": 234421394}, "PAN-X.ta": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.ta", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 277625, "num_examples": 1000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 278114, "num_examples": 1000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 4122130, "num_examples": 15000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 4677869, "size_in_bytes": 238686753}, "PAN-X.te": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.te", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 293281, "num_examples": 1000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 296963, "num_examples": 1000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 295410, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 885654, "size_in_bytes": 234894538}, "PAN-X.th": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.th", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 13262737, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 13586928, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 27133029, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 53982694, "size_in_bytes": 287991578}, "PAN-X.tl": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.tl", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 114156, "num_examples": 1000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 117904, "num_examples": 1000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 1168717, "num_examples": 10000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 1400777, "size_in_bytes": 235409661}, "PAN-X.tr": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.tr", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 1915352, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 1911503, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 3779170, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 7606025, "size_in_bytes": 241614909}, "PAN-X.ur": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.ur", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 152148, "num_examples": 1000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 151922, "num_examples": 1000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 3072276, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 3376346, "size_in_bytes": 237385230}, "PAN-X.vi": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.vi", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 1565143, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 1580216, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 3153227, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 6298586, "size_in_bytes": 240307470}, "PAN-X.yo": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.yo", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 13245, "num_examples": 100, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 13533, "num_examples": 100, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 14709, "num_examples": 100, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 41487, "size_in_bytes": 234050371}, "PAN-X.zh": {"description": "The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been\nconstructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset\ncan be loaded with the DaNLP package:\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{pan-x,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Xiaoman, Pan and Boliang, Zhang and Jonathan, May and Joel, Nothman and Kevin, Knight and Heng, Ji},\n volume={Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}\n year={2017}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/afshinrahimi/mmner", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ner_tags": {"feature": {"num_classes": 7, "names": ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}, "langs": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAN-X.zh", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 4491325, "num_examples": 10000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 4363172, "num_examples": 10000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 8832051, "num_examples": 20000, "dataset_name": "xtreme"}}, "download_checksums": {"https://s3.amazonaws.com/datasets.huggingface.co/wikiann/1.1.0/panx_dataset.zip": {"num_bytes": 234008884, "checksum": "164814b64f749ad000988b5ab45050b5116913b3db466ab173008955ddf649a4"}}, "download_size": 234008884, "post_processing_size": null, "dataset_size": 17686548, "size_in_bytes": 251695432}, "MLQA.ar.ar": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.ar.ar", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 8368114, "num_examples": 5335, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 824108, "num_examples": 517, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 9192222, "size_in_bytes": 84911272}, "MLQA.ar.de": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.ar.de", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 2183942, "num_examples": 1649, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 364837, "num_examples": 207, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 2548779, "size_in_bytes": 78267829}, "MLQA.ar.vi": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.ar.vi", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 3290629, "num_examples": 2047, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 288446, "num_examples": 163, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 3579075, "size_in_bytes": 79298125}, "MLQA.ar.zh": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.ar.zh", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 3229872, "num_examples": 1912, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 340049, "num_examples": 188, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 3569921, "size_in_bytes": 79288971}, "MLQA.ar.en": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.ar.en", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 8225662, "num_examples": 5335, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 810089, "num_examples": 517, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 9035751, "size_in_bytes": 84754801}, "MLQA.ar.es": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.ar.es", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 3041378, "num_examples": 1978, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 228180, "num_examples": 161, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 3269558, "size_in_bytes": 78988608}, "MLQA.ar.hi": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.ar.hi", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 3039396, "num_examples": 1831, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 281770, "num_examples": 186, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 3321166, "size_in_bytes": 79040216}, "MLQA.de.ar": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.de.ar", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 1620006, "num_examples": 1649, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 200174, "num_examples": 207, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 1820180, "size_in_bytes": 77539230}, "MLQA.de.de": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.de.de", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 4366102, "num_examples": 4517, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 488367, "num_examples": 512, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 4854469, "size_in_bytes": 80573519}, "MLQA.de.vi": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.de.vi", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 1688483, "num_examples": 1675, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 216075, "num_examples": 182, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 1904558, "size_in_bytes": 77623608}, "MLQA.de.zh": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.de.zh", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 1679180, "num_examples": 1621, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 184318, "num_examples": 190, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 1863498, "size_in_bytes": 77582548}, "MLQA.de.en": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.de.en", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 4343144, "num_examples": 4517, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 485894, "num_examples": 512, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 4829038, "size_in_bytes": 80548088}, "MLQA.de.es": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.de.es", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 1716615, "num_examples": 1776, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 170582, "num_examples": 196, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 1887197, "size_in_bytes": 77606247}, "MLQA.de.hi": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.de.hi", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 1371074, "num_examples": 1430, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 153871, "num_examples": 163, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 1524945, "size_in_bytes": 77243995}, "MLQA.vi.ar": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.vi.ar", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 3205185, "num_examples": 2047, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 230335, "num_examples": 163, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 3435520, "size_in_bytes": 79154570}, "MLQA.vi.de": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.vi.de", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 2227033, "num_examples": 1675, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 277185, "num_examples": 182, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 2504218, "size_in_bytes": 78223268}, "MLQA.vi.vi": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.vi.vi", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 7922085, "num_examples": 5495, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 726518, "num_examples": 511, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 8648603, "size_in_bytes": 84367653}, "MLQA.vi.zh": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.vi.zh", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 2989660, "num_examples": 1943, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 269389, "num_examples": 184, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 3259049, "size_in_bytes": 78978099}, "MLQA.vi.en": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.vi.en", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 7843431, "num_examples": 5495, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 719273, "num_examples": 511, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 8562704, "size_in_bytes": 84281754}, "MLQA.vi.es": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.vi.es", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 2866597, "num_examples": 2018, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 283461, "num_examples": 189, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 3150058, "size_in_bytes": 78869108}, "MLQA.vi.hi": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.vi.hi", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 2776664, "num_examples": 1947, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 255007, "num_examples": 177, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 3031671, "size_in_bytes": 78750721}, "MLQA.zh.ar": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.zh.ar", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 1731483, "num_examples": 1912, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 175349, "num_examples": 188, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 1906832, "size_in_bytes": 77625882}, "MLQA.zh.de": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.zh.de", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 1390018, "num_examples": 1621, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 174605, "num_examples": 190, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 1564623, "size_in_bytes": 77283673}, "MLQA.zh.vi": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.zh.vi", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 1806186, "num_examples": 1943, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 172934, "num_examples": 184, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 1979120, "size_in_bytes": 77698170}, "MLQA.zh.zh": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.zh.zh", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 4422350, "num_examples": 5137, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 443810, "num_examples": 504, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 4866160, "size_in_bytes": 80585210}, "MLQA.zh.en": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.zh.en", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 4450985, "num_examples": 5137, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 446868, "num_examples": 504, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 4897853, "size_in_bytes": 80616903}, "MLQA.zh.es": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.zh.es", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 1736283, "num_examples": 1947, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 138073, "num_examples": 161, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 1874356, "size_in_bytes": 77593406}, "MLQA.zh.hi": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.zh.hi", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 1578219, "num_examples": 1767, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 184401, "num_examples": 189, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 1762620, "size_in_bytes": 77481670}, "MLQA.en.ar": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.en.ar", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 6739219, "num_examples": 5335, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 630843, "num_examples": 517, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 7370062, "size_in_bytes": 83089112}, "MLQA.en.de": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.en.de", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 5056722, "num_examples": 4517, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 594936, "num_examples": 512, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 5651658, "size_in_bytes": 81370708}, "MLQA.en.vi": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.en.vi", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 7056698, "num_examples": 5495, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 640646, "num_examples": 511, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 7697344, "size_in_bytes": 83416394}, "MLQA.en.zh": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.en.zh", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 6539307, "num_examples": 5137, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 608444, "num_examples": 504, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 7147751, "size_in_bytes": 82866801}, "MLQA.en.en": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.en.en", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 14004648, "num_examples": 11590, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 1329112, "num_examples": 1148, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 15333760, "size_in_bytes": 91052810}, "MLQA.en.es": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.en.es", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 6179249, "num_examples": 5253, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 555462, "num_examples": 500, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 6734711, "size_in_bytes": 82453761}, "MLQA.en.hi": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.en.hi", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 6378866, "num_examples": 4918, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 623171, "num_examples": 507, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 7002037, "size_in_bytes": 82721087}, "MLQA.es.ar": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.es.ar", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 1740282, "num_examples": 1978, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 148649, "num_examples": 161, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 1888931, "size_in_bytes": 77607981}, "MLQA.es.de": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.es.de", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 1404025, "num_examples": 1776, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 144186, "num_examples": 196, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 1548211, "size_in_bytes": 77267261}, "MLQA.es.vi": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.es.vi", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 1747969, "num_examples": 2018, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 176841, "num_examples": 189, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 1924810, "size_in_bytes": 77643860}, "MLQA.es.zh": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.es.zh", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 1678451, "num_examples": 1947, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 126646, "num_examples": 161, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 1805097, "size_in_bytes": 77524147}, "MLQA.es.en": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.es.en", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 4362737, "num_examples": 5253, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 419068, "num_examples": 500, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 4781805, "size_in_bytes": 80500855}, "MLQA.es.es": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.es.es", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 4394333, "num_examples": 5253, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 422071, "num_examples": 500, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 4816404, "size_in_bytes": 80535454}, "MLQA.es.hi": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.es.hi", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 1523523, "num_examples": 1723, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 181834, "num_examples": 187, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 1705357, "size_in_bytes": 77424407}, "MLQA.hi.ar": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.hi.ar", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 4445589, "num_examples": 1831, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 410424, "num_examples": 186, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 4856013, "size_in_bytes": 80575063}, "MLQA.hi.de": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.hi.de", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 3022864, "num_examples": 1430, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 301713, "num_examples": 163, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 3324577, "size_in_bytes": 79043627}, "MLQA.hi.vi": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.hi.vi", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 4743484, "num_examples": 1947, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 419106, "num_examples": 177, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 5162590, "size_in_bytes": 80881640}, "MLQA.hi.zh": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.hi.zh", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 4354875, "num_examples": 1767, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 424246, "num_examples": 189, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 4779121, "size_in_bytes": 80498171}, "MLQA.hi.en": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.hi.en", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 11449261, "num_examples": 4918, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 1097857, "num_examples": 507, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 12547118, "size_in_bytes": 88266168}, "MLQA.hi.es": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.hi.es", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 3862229, "num_examples": 1723, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 420402, "num_examples": 187, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 4282631, "size_in_bytes": 80001681}, "MLQA.hi.hi": {"description": " MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.\nMLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,\nGerman, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between\n4 different languages on average.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "@article{lewis2019mlqa,\ntitle={MLQA: Evaluating Cross-lingual Extractive Question Answering},\nauthor={Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},\njournal={arXiv preprint arXiv:1910.07475},\nyear={2019}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/MLQA", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "MLQA.hi.hi", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 11810475, "num_examples": 4918, "dataset_name": "xtreme"}, "validation": {"name": "validation", "num_bytes": 1136784, "num_examples": 507, "dataset_name": "xtreme"}}, "download_checksums": {"https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip": {"num_bytes": 75719050, "checksum": "246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8"}}, "download_size": 75719050, "post_processing_size": null, "dataset_size": 12947259, "size_in_bytes": 88666309}, "XQuAD.ar": {"description": "XQuAD (Cross-lingual Question Answering Dataset) is a benchmark dataset for evaluating cross-lingual question\nanswering performance. The dataset consists of a subset of 240 paragraphs and 1190 question-answer pairs from\nthe development set of SQuAD v1.1 (Rajpurkar et al., 2016) together with their professional translations into\nten languages: Spanish, German, Greek, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, and Hindi. Consequently,\nthe dataset is entirely parallel across 11 languages.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n @article{Artetxe:etal:2019,\n author = {Mikel Artetxe and Sebastian Ruder and Dani Yogatama},\n title = {On the cross-lingual transferability of monolingual representations},\n journal = {CoRR},\n volume = {abs/1910.11856},\n year = {2019},\n archivePrefix = {arXiv},\n eprint = {1910.11856}\n}\n\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/deepmind/xquad", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "XQuAD.ar", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 1722799, "num_examples": 1190, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/deepmind/xquad/raw/master/xquad.ar.json": {"num_bytes": 1582988, "checksum": "abdabd7afed5c635d99cca0f3f0d0c9d9ed0bc77451e963c2e4e0638c29e486d"}}, "download_size": 1582988, "post_processing_size": null, "dataset_size": 1722799, "size_in_bytes": 3305787}, "XQuAD.de": {"description": "XQuAD (Cross-lingual Question Answering Dataset) is a benchmark dataset for evaluating cross-lingual question\nanswering performance. The dataset consists of a subset of 240 paragraphs and 1190 question-answer pairs from\nthe development set of SQuAD v1.1 (Rajpurkar et al., 2016) together with their professional translations into\nten languages: Spanish, German, Greek, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, and Hindi. Consequently,\nthe dataset is entirely parallel across 11 languages.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n @article{Artetxe:etal:2019,\n author = {Mikel Artetxe and Sebastian Ruder and Dani Yogatama},\n title = {On the cross-lingual transferability of monolingual representations},\n journal = {CoRR},\n volume = {abs/1910.11856},\n year = {2019},\n archivePrefix = {arXiv},\n eprint = {1910.11856}\n}\n\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/deepmind/xquad", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "XQuAD.de", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 1283301, "num_examples": 1190, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/deepmind/xquad/raw/master/xquad.de.json": {"num_bytes": 669810, "checksum": "990b5d746746ed65ed4702ea5f35f99ffa4e2f1c390c07d003642acd937916f9"}}, "download_size": 669810, "post_processing_size": null, "dataset_size": 1283301, "size_in_bytes": 1953111}, "XQuAD.vi": {"description": "XQuAD (Cross-lingual Question Answering Dataset) is a benchmark dataset for evaluating cross-lingual question\nanswering performance. The dataset consists of a subset of 240 paragraphs and 1190 question-answer pairs from\nthe development set of SQuAD v1.1 (Rajpurkar et al., 2016) together with their professional translations into\nten languages: Spanish, German, Greek, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, and Hindi. Consequently,\nthe dataset is entirely parallel across 11 languages.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n @article{Artetxe:etal:2019,\n author = {Mikel Artetxe and Sebastian Ruder and Dani Yogatama},\n title = {On the cross-lingual transferability of monolingual representations},\n journal = {CoRR},\n volume = {abs/1910.11856},\n year = {2019},\n archivePrefix = {arXiv},\n eprint = {1910.11856}\n}\n\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/deepmind/xquad", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "XQuAD.vi", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 1477239, "num_examples": 1190, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/deepmind/xquad/raw/master/xquad.vi.json": {"num_bytes": 911401, "checksum": "f619a1eb11fb42d3ab0834259e488a65f585447ef6154437bfb7199d85161a04"}}, "download_size": 911401, "post_processing_size": null, "dataset_size": 1477239, "size_in_bytes": 2388640}, "XQuAD.zh": {"description": "XQuAD (Cross-lingual Question Answering Dataset) is a benchmark dataset for evaluating cross-lingual question\nanswering performance. The dataset consists of a subset of 240 paragraphs and 1190 question-answer pairs from\nthe development set of SQuAD v1.1 (Rajpurkar et al., 2016) together with their professional translations into\nten languages: Spanish, German, Greek, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, and Hindi. Consequently,\nthe dataset is entirely parallel across 11 languages.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n @article{Artetxe:etal:2019,\n author = {Mikel Artetxe and Sebastian Ruder and Dani Yogatama},\n title = {On the cross-lingual transferability of monolingual representations},\n journal = {CoRR},\n volume = {abs/1910.11856},\n year = {2019},\n archivePrefix = {arXiv},\n eprint = {1910.11856}\n}\n\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/deepmind/xquad", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "XQuAD.zh", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 984241, "num_examples": 1190, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/deepmind/xquad/raw/master/xquad.zh.json": {"num_bytes": 808652, "checksum": "691d0b3359bc6b8faa8de931dfdfe21d50a65861ae348e32a0d1a0190b0c8835"}}, "download_size": 808652, "post_processing_size": null, "dataset_size": 984241, "size_in_bytes": 1792893}, "XQuAD.en": {"description": "XQuAD (Cross-lingual Question Answering Dataset) is a benchmark dataset for evaluating cross-lingual question\nanswering performance. The dataset consists of a subset of 240 paragraphs and 1190 question-answer pairs from\nthe development set of SQuAD v1.1 (Rajpurkar et al., 2016) together with their professional translations into\nten languages: Spanish, German, Greek, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, and Hindi. Consequently,\nthe dataset is entirely parallel across 11 languages.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n @article{Artetxe:etal:2019,\n author = {Mikel Artetxe and Sebastian Ruder and Dani Yogatama},\n title = {On the cross-lingual transferability of monolingual representations},\n journal = {CoRR},\n volume = {abs/1910.11856},\n year = {2019},\n archivePrefix = {arXiv},\n eprint = {1910.11856}\n}\n\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/deepmind/xquad", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "XQuAD.en", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 1116123, "num_examples": 1190, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/deepmind/xquad/raw/master/xquad.en.json": {"num_bytes": 609383, "checksum": "e4c57d1c9143aaa1c5d265ba5987a65f4e69528d2a98f29d6e75019b10344f29"}}, "download_size": 609383, "post_processing_size": null, "dataset_size": 1116123, "size_in_bytes": 1725506}, "XQuAD.es": {"description": "XQuAD (Cross-lingual Question Answering Dataset) is a benchmark dataset for evaluating cross-lingual question\nanswering performance. The dataset consists of a subset of 240 paragraphs and 1190 question-answer pairs from\nthe development set of SQuAD v1.1 (Rajpurkar et al., 2016) together with their professional translations into\nten languages: Spanish, German, Greek, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, and Hindi. Consequently,\nthe dataset is entirely parallel across 11 languages.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n @article{Artetxe:etal:2019,\n author = {Mikel Artetxe and Sebastian Ruder and Dani Yogatama},\n title = {On the cross-lingual transferability of monolingual representations},\n journal = {CoRR},\n volume = {abs/1910.11856},\n year = {2019},\n archivePrefix = {arXiv},\n eprint = {1910.11856}\n}\n\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/deepmind/xquad", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "XQuAD.es", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 1273499, "num_examples": 1190, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/deepmind/xquad/raw/master/xquad.es.json": {"num_bytes": 684322, "checksum": "dcbae93ec3a9f4b9e78fd834a171d6f96c1a875e10e15b7530b7e4ef4971e37e"}}, "download_size": 684322, "post_processing_size": null, "dataset_size": 1273499, "size_in_bytes": 1957821}, "XQuAD.hi": {"description": "XQuAD (Cross-lingual Question Answering Dataset) is a benchmark dataset for evaluating cross-lingual question\nanswering performance. The dataset consists of a subset of 240 paragraphs and 1190 question-answer pairs from\nthe development set of SQuAD v1.1 (Rajpurkar et al., 2016) together with their professional translations into\nten languages: Spanish, German, Greek, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, and Hindi. Consequently,\nthe dataset is entirely parallel across 11 languages.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n @article{Artetxe:etal:2019,\n author = {Mikel Artetxe and Sebastian Ruder and Dani Yogatama},\n title = {On the cross-lingual transferability of monolingual representations},\n journal = {CoRR},\n volume = {abs/1910.11856},\n year = {2019},\n archivePrefix = {arXiv},\n eprint = {1910.11856}\n}\n\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/deepmind/xquad", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "XQuAD.hi", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 2682975, "num_examples": 1190, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/deepmind/xquad/raw/master/xquad.hi.json": {"num_bytes": 1680538, "checksum": "df2cce3532b37e9beb8979704b5c9a4bf874358f105395a298b89427b43b9d24"}}, "download_size": 1680538, "post_processing_size": null, "dataset_size": 2682975, "size_in_bytes": 4363513}, "XQuAD.el": {"description": "XQuAD (Cross-lingual Question Answering Dataset) is a benchmark dataset for evaluating cross-lingual question\nanswering performance. The dataset consists of a subset of 240 paragraphs and 1190 question-answer pairs from\nthe development set of SQuAD v1.1 (Rajpurkar et al., 2016) together with their professional translations into\nten languages: Spanish, German, Greek, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, and Hindi. Consequently,\nthe dataset is entirely parallel across 11 languages.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n @article{Artetxe:etal:2019,\n author = {Mikel Artetxe and Sebastian Ruder and Dani Yogatama},\n title = {On the cross-lingual transferability of monolingual representations},\n journal = {CoRR},\n volume = {abs/1910.11856},\n year = {2019},\n archivePrefix = {arXiv},\n eprint = {1910.11856}\n}\n\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/deepmind/xquad", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "XQuAD.el", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 2206690, "num_examples": 1190, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/deepmind/xquad/raw/master/xquad.el.json": {"num_bytes": 1918889, "checksum": "821cf0f88e73fa258fd2f548b19b6ec39f7025059e16f6f9fc8cd797c9c3663e"}}, "download_size": 1918889, "post_processing_size": null, "dataset_size": 2206690, "size_in_bytes": 4125579}, "XQuAD.ru": {"description": "XQuAD (Cross-lingual Question Answering Dataset) is a benchmark dataset for evaluating cross-lingual question\nanswering performance. The dataset consists of a subset of 240 paragraphs and 1190 question-answer pairs from\nthe development set of SQuAD v1.1 (Rajpurkar et al., 2016) together with their professional translations into\nten languages: Spanish, German, Greek, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, and Hindi. Consequently,\nthe dataset is entirely parallel across 11 languages.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n @article{Artetxe:etal:2019,\n author = {Mikel Artetxe and Sebastian Ruder and Dani Yogatama},\n title = {On the cross-lingual transferability of monolingual representations},\n journal = {CoRR},\n volume = {abs/1910.11856},\n year = {2019},\n archivePrefix = {arXiv},\n eprint = {1910.11856}\n}\n\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/deepmind/xquad", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "XQuAD.ru", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 2136990, "num_examples": 1190, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/deepmind/xquad/raw/master/xquad.ru.json": {"num_bytes": 1896368, "checksum": "208d5b1aa154c52b1b5c5eda16281e455e8fd198cdb9af3f469f0d6037d973bf"}}, "download_size": 1896368, "post_processing_size": null, "dataset_size": 2136990, "size_in_bytes": 4033358}, "XQuAD.th": {"description": "XQuAD (Cross-lingual Question Answering Dataset) is a benchmark dataset for evaluating cross-lingual question\nanswering performance. The dataset consists of a subset of 240 paragraphs and 1190 question-answer pairs from\nthe development set of SQuAD v1.1 (Rajpurkar et al., 2016) together with their professional translations into\nten languages: Spanish, German, Greek, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, and Hindi. Consequently,\nthe dataset is entirely parallel across 11 languages.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n @article{Artetxe:etal:2019,\n author = {Mikel Artetxe and Sebastian Ruder and Dani Yogatama},\n title = {On the cross-lingual transferability of monolingual representations},\n journal = {CoRR},\n volume = {abs/1910.11856},\n year = {2019},\n archivePrefix = {arXiv},\n eprint = {1910.11856}\n}\n\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/deepmind/xquad", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "XQuAD.th", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 2854959, "num_examples": 1190, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/deepmind/xquad/raw/master/xquad.th.json": {"num_bytes": 1809143, "checksum": "5cdda11d0e1e075f7872abf4e6ae830388ce7f617964d542308e9ae4257e0f43"}}, "download_size": 1809143, "post_processing_size": null, "dataset_size": 2854959, "size_in_bytes": 4664102}, "XQuAD.tr": {"description": "XQuAD (Cross-lingual Question Answering Dataset) is a benchmark dataset for evaluating cross-lingual question\nanswering performance. The dataset consists of a subset of 240 paragraphs and 1190 question-answer pairs from\nthe development set of SQuAD v1.1 (Rajpurkar et al., 2016) together with their professional translations into\nten languages: Spanish, German, Greek, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, and Hindi. Consequently,\nthe dataset is entirely parallel across 11 languages.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n @article{Artetxe:etal:2019,\n author = {Mikel Artetxe and Sebastian Ruder and Dani Yogatama},\n title = {On the cross-lingual transferability of monolingual representations},\n journal = {CoRR},\n volume = {abs/1910.11856},\n year = {2019},\n archivePrefix = {arXiv},\n eprint = {1910.11856}\n}\n\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/deepmind/xquad", "license": "", "features": {"id": {"dtype": "string", "id": null, "_type": "Value"}, "context": {"dtype": "string", "id": null, "_type": "Value"}, "question": {"dtype": "string", "id": null, "_type": "Value"}, "answers": {"feature": {"answer_start": {"dtype": "int32", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "XQuAD.tr", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 1210763, "num_examples": 1190, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/deepmind/xquad/raw/master/xquad.tr.json": {"num_bytes": 729506, "checksum": "92179a564774b7696100d144c1e10870d0a966b6fccbdd254a65b9d2ab1971cc"}}, "download_size": 729506, "post_processing_size": null, "dataset_size": 1210763, "size_in_bytes": 1940269}, "bucc18.de": {"description": "Building and Using Comparable Corpora\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://comparable.limsi.fr/bucc2018/", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "bucc18.de", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 248707, "num_examples": 1038, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 2325701, "num_examples": 9580, "dataset_name": "xtreme"}}, "download_checksums": {"https://comparable.limsi.fr/bucc2018/bucc2018-de-en.training-gold.tar.bz2": {"num_bytes": 28189548, "checksum": "766e0fdebbd1438fb87c21254828eb13c8b997d8fbab002103dd060dcac50c5c"}, "https://comparable.limsi.fr/bucc2018/bucc2018-de-en.sample-gold.tar.bz2": {"num_bytes": 2529652, "checksum": "8e16ba8b02ef8d648b06adfbd6dfb188f43524e18f97b2b12a14a086caac62f3"}}, "download_size": 30719200, "post_processing_size": null, "dataset_size": 2574408, "size_in_bytes": 33293608}, "bucc18.fr": {"description": "Building and Using Comparable Corpora\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://comparable.limsi.fr/bucc2018/", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "bucc18.fr", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 212513, "num_examples": 929, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 2082419, "num_examples": 9086, "dataset_name": "xtreme"}}, "download_checksums": {"https://comparable.limsi.fr/bucc2018/bucc2018-fr-en.training-gold.tar.bz2": {"num_bytes": 20757713, "checksum": "53c0d0e7dd97dc89593a2db25a26c5f0ccdc7113e8451263bb0c80e4c4c7dc30"}, "https://comparable.limsi.fr/bucc2018/bucc2018-fr-en.sample-gold.tar.bz2": {"num_bytes": 1948831, "checksum": "df9eb3966954e163c9264076f7c9c1eb56d9d8a91855f9d3afbf2c0fdaef0a08"}}, "download_size": 22706544, "post_processing_size": null, "dataset_size": 2294932, "size_in_bytes": 25001476}, "bucc18.zh": {"description": "Building and Using Comparable Corpora\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://comparable.limsi.fr/bucc2018/", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "bucc18.zh", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 55739, "num_examples": 257, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 415925, "num_examples": 1899, "dataset_name": "xtreme"}}, "download_checksums": {"https://comparable.limsi.fr/bucc2018/bucc2018-zh-en.training-gold.tar.bz2": {"num_bytes": 6344925, "checksum": "3facb71798277c8f44dc78c1f8ae2110f254d0e14799f3508eedd54b4236877a"}, "https://comparable.limsi.fr/bucc2018/bucc2018-zh-en.sample-gold.tar.bz2": {"num_bytes": 769869, "checksum": "a3425be5c0320ee131a0927b66c3e29befb3b481ebf1b87257e660e514bc16ac"}}, "download_size": 7114794, "post_processing_size": null, "dataset_size": 471664, "size_in_bytes": 7586458}, "bucc18.ru": {"description": "Building and Using Comparable Corpora\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://comparable.limsi.fr/bucc2018/", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "bucc18.ru", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 761347, "num_examples": 2374, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 4641678, "num_examples": 14435, "dataset_name": "xtreme"}}, "download_checksums": {"https://comparable.limsi.fr/bucc2018/bucc2018-ru-en.training-gold.tar.bz2": {"num_bytes": 37085079, "checksum": "1895df56e936ca3d4f5b12299ceffe0b7ff4806584c40bdaa3ae1d445f25afa5"}, "https://comparable.limsi.fr/bucc2018/bucc2018-ru-en.sample-gold.tar.bz2": {"num_bytes": 4269233, "checksum": "fce3cabc7ee50ddb4b18aa6fb090e2669c8383d2a29fc97eed6ae70fed9a23e5"}}, "download_size": 41354312, "post_processing_size": null, "dataset_size": 5403025, "size_in_bytes": 46757337}, "PAWS-X.de": {"description": "\nThis dataset contains 23,659 human translated PAWS evaluation pairs and 296,406 machine translated training\npairs in six typologically distinct languages: French, Spanish, German, Chinese, Japanese, and Korean. All\ntranslated pairs are sourced from examples in PAWS-Wiki.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @InProceedings{pawsx2019emnlp,\n title = {{PAWS-X: A Cross-lingual Adversarial Dataset for Paraphrase Identification}},\n author = {Yang, Yinfei and Zhang, Yuan and Tar, Chris and Baldridge, Jason},\n booktitle = {Proc. of EMNLP},\n year = {2019}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/google-research-datasets/paws/tree/master/pawsx", "license": "", "features": {"sentence1": {"dtype": "string", "id": null, "_type": "Value"}, "sentence2": {"dtype": "string", "id": null, "_type": "Value"}, "label": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAWS-X.de", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 500009, "num_examples": 2000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 510194, "num_examples": 2000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 12451883, "num_examples": 49380, "dataset_name": "xtreme"}}, "download_checksums": {"https://storage.googleapis.com/paws/pawsx/x-final.tar.gz": {"num_bytes": 30282057, "checksum": "4146db499101d66e68ae4c8ed3cf9dadecd625f44b7d8cf3d4a0fe93afc4fd9f"}}, "download_size": 30282057, "post_processing_size": null, "dataset_size": 13462086, "size_in_bytes": 43744143}, "PAWS-X.en": {"description": "\nThis dataset contains 23,659 human translated PAWS evaluation pairs and 296,406 machine translated training\npairs in six typologically distinct languages: French, Spanish, German, Chinese, Japanese, and Korean. All\ntranslated pairs are sourced from examples in PAWS-Wiki.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @InProceedings{pawsx2019emnlp,\n title = {{PAWS-X: A Cross-lingual Adversarial Dataset for Paraphrase Identification}},\n author = {Yang, Yinfei and Zhang, Yuan and Tar, Chris and Baldridge, Jason},\n booktitle = {Proc. of EMNLP},\n year = {2019}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/google-research-datasets/paws/tree/master/pawsx", "license": "", "features": {"sentence1": {"dtype": "string", "id": null, "_type": "Value"}, "sentence2": {"dtype": "string", "id": null, "_type": "Value"}, "label": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAWS-X.en", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 478291, "num_examples": 2000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 480738, "num_examples": 2000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 11827719, "num_examples": 49175, "dataset_name": "xtreme"}}, "download_checksums": {"https://storage.googleapis.com/paws/pawsx/x-final.tar.gz": {"num_bytes": 30282057, "checksum": "4146db499101d66e68ae4c8ed3cf9dadecd625f44b7d8cf3d4a0fe93afc4fd9f"}}, "download_size": 30282057, "post_processing_size": null, "dataset_size": 12786748, "size_in_bytes": 43068805}, "PAWS-X.es": {"description": "\nThis dataset contains 23,659 human translated PAWS evaluation pairs and 296,406 machine translated training\npairs in six typologically distinct languages: French, Spanish, German, Chinese, Japanese, and Korean. All\ntranslated pairs are sourced from examples in PAWS-Wiki.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @InProceedings{pawsx2019emnlp,\n title = {{PAWS-X: A Cross-lingual Adversarial Dataset for Paraphrase Identification}},\n author = {Yang, Yinfei and Zhang, Yuan and Tar, Chris and Baldridge, Jason},\n booktitle = {Proc. of EMNLP},\n year = {2019}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/google-research-datasets/paws/tree/master/pawsx", "license": "", "features": {"sentence1": {"dtype": "string", "id": null, "_type": "Value"}, "sentence2": {"dtype": "string", "id": null, "_type": "Value"}, "label": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAWS-X.es", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 494069, "num_examples": 1961, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 505047, "num_examples": 2000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 12462107, "num_examples": 49401, "dataset_name": "xtreme"}}, "download_checksums": {"https://storage.googleapis.com/paws/pawsx/x-final.tar.gz": {"num_bytes": 30282057, "checksum": "4146db499101d66e68ae4c8ed3cf9dadecd625f44b7d8cf3d4a0fe93afc4fd9f"}}, "download_size": 30282057, "post_processing_size": null, "dataset_size": 13461223, "size_in_bytes": 43743280}, "PAWS-X.fr": {"description": "\nThis dataset contains 23,659 human translated PAWS evaluation pairs and 296,406 machine translated training\npairs in six typologically distinct languages: French, Spanish, German, Chinese, Japanese, and Korean. All\ntranslated pairs are sourced from examples in PAWS-Wiki.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @InProceedings{pawsx2019emnlp,\n title = {{PAWS-X: A Cross-lingual Adversarial Dataset for Paraphrase Identification}},\n author = {Yang, Yinfei and Zhang, Yuan and Tar, Chris and Baldridge, Jason},\n booktitle = {Proc. of EMNLP},\n year = {2019}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/google-research-datasets/paws/tree/master/pawsx", "license": "", "features": {"sentence1": {"dtype": "string", "id": null, "_type": "Value"}, "sentence2": {"dtype": "string", "id": null, "_type": "Value"}, "label": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAWS-X.fr", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 516111, "num_examples": 1988, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 521031, "num_examples": 2000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 12948512, "num_examples": 49399, "dataset_name": "xtreme"}}, "download_checksums": {"https://storage.googleapis.com/paws/pawsx/x-final.tar.gz": {"num_bytes": 30282057, "checksum": "4146db499101d66e68ae4c8ed3cf9dadecd625f44b7d8cf3d4a0fe93afc4fd9f"}}, "download_size": 30282057, "post_processing_size": null, "dataset_size": 13985654, "size_in_bytes": 44267711}, "PAWS-X.ja": {"description": "\nThis dataset contains 23,659 human translated PAWS evaluation pairs and 296,406 machine translated training\npairs in six typologically distinct languages: French, Spanish, German, Chinese, Japanese, and Korean. All\ntranslated pairs are sourced from examples in PAWS-Wiki.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @InProceedings{pawsx2019emnlp,\n title = {{PAWS-X: A Cross-lingual Adversarial Dataset for Paraphrase Identification}},\n author = {Yang, Yinfei and Zhang, Yuan and Tar, Chris and Baldridge, Jason},\n booktitle = {Proc. of EMNLP},\n year = {2019}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/google-research-datasets/paws/tree/master/pawsx", "license": "", "features": {"sentence1": {"dtype": "string", "id": null, "_type": "Value"}, "sentence2": {"dtype": "string", "id": null, "_type": "Value"}, "label": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAWS-X.ja", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 647774, "num_examples": 2000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 654640, "num_examples": 2000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 14695653, "num_examples": 49401, "dataset_name": "xtreme"}}, "download_checksums": {"https://storage.googleapis.com/paws/pawsx/x-final.tar.gz": {"num_bytes": 30282057, "checksum": "4146db499101d66e68ae4c8ed3cf9dadecd625f44b7d8cf3d4a0fe93afc4fd9f"}}, "download_size": 30282057, "post_processing_size": null, "dataset_size": 15998067, "size_in_bytes": 46280124}, "PAWS-X.ko": {"description": "\nThis dataset contains 23,659 human translated PAWS evaluation pairs and 296,406 machine translated training\npairs in six typologically distinct languages: French, Spanish, German, Chinese, Japanese, and Korean. All\ntranslated pairs are sourced from examples in PAWS-Wiki.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @InProceedings{pawsx2019emnlp,\n title = {{PAWS-X: A Cross-lingual Adversarial Dataset for Paraphrase Identification}},\n author = {Yang, Yinfei and Zhang, Yuan and Tar, Chris and Baldridge, Jason},\n booktitle = {Proc. of EMNLP},\n year = {2019}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/google-research-datasets/paws/tree/master/pawsx", "license": "", "features": {"sentence1": {"dtype": "string", "id": null, "_type": "Value"}, "sentence2": {"dtype": "string", "id": null, "_type": "Value"}, "label": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAWS-X.ko", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 540787, "num_examples": 2000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 547978, "num_examples": 1999, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 13542657, "num_examples": 49164, "dataset_name": "xtreme"}}, "download_checksums": {"https://storage.googleapis.com/paws/pawsx/x-final.tar.gz": {"num_bytes": 30282057, "checksum": "4146db499101d66e68ae4c8ed3cf9dadecd625f44b7d8cf3d4a0fe93afc4fd9f"}}, "download_size": 30282057, "post_processing_size": null, "dataset_size": 14631422, "size_in_bytes": 44913479}, "PAWS-X.zh": {"description": "\nThis dataset contains 23,659 human translated PAWS evaluation pairs and 296,406 machine translated training\npairs in six typologically distinct languages: French, Spanish, German, Chinese, Japanese, and Korean. All\ntranslated pairs are sourced from examples in PAWS-Wiki.\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @InProceedings{pawsx2019emnlp,\n title = {{PAWS-X: A Cross-lingual Adversarial Dataset for Paraphrase Identification}},\n author = {Yang, Yinfei and Zhang, Yuan and Tar, Chris and Baldridge, Jason},\n booktitle = {Proc. of EMNLP},\n year = {2019}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/google-research-datasets/paws/tree/master/pawsx", "license": "", "features": {"sentence1": {"dtype": "string", "id": null, "_type": "Value"}, "sentence2": {"dtype": "string", "id": null, "_type": "Value"}, "label": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "PAWS-X.zh", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 459120, "num_examples": 2000, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 460638, "num_examples": 2000, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 10469712, "num_examples": 49401, "dataset_name": "xtreme"}}, "download_checksums": {"https://storage.googleapis.com/paws/pawsx/x-final.tar.gz": {"num_bytes": 30282057, "checksum": "4146db499101d66e68ae4c8ed3cf9dadecd625f44b7d8cf3d4a0fe93afc4fd9f"}}, "download_size": 30282057, "post_processing_size": null, "dataset_size": 11389470, "size_in_bytes": 41671527}, "tatoeba.afr": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.afr", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 179651, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.afr-eng.afr": {"num_bytes": 30586, "checksum": "7bb9e073ad8422d6bfdec7c9ebdcef8ac486e72b237200e447923a8b921a0a56"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.afr-eng.eng": {"num_bytes": 29049, "checksum": "0b700c125efb4030b4cc50d9d765d5884afc24f39296f29b028a1b2a8512034f"}}, "download_size": 59635, "post_processing_size": null, "dataset_size": 179651, "size_in_bytes": 239286}, "tatoeba.ara": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.ara", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 192666, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.ara-eng.ara": {"num_bytes": 43582, "checksum": "e67a0ae072b79cd9e8eb09f166c3bc0b23488d39f5720f2ee0a8350ae17b719f"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.ara-eng.eng": {"num_bytes": 29068, "checksum": "3c11838c963f598a52dcf2f452b666353538257001db5c59c3a5f54a999b336b"}}, "download_size": 72650, "post_processing_size": null, "dataset_size": 192666, "size_in_bytes": 265316}, "tatoeba.ben": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.ben", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 211719, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.ben-eng.ben": {"num_bytes": 65990, "checksum": "ac3385695d6a6c7e5d18e38ad4b8b7d3780f3df23dd0ff3f539071b8269a8613"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.ben-eng.eng": {"num_bytes": 25713, "checksum": "67fbe75fec549d436c3356b6d6f8dd53179b6a908661b5d507d28c7fee83350e"}}, "download_size": 91703, "post_processing_size": null, "dataset_size": 211719, "size_in_bytes": 303422}, "tatoeba.bul": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.bul", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 222295, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.bul-eng.bul": {"num_bytes": 65500, "checksum": "f9fa90cf3599d8c87f7a6ed22f5d648e3ce6687c705656a8c8ea088d891f79d5"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.bul-eng.eng": {"num_bytes": 36779, "checksum": "4b03b3b52ffb7cf4286e0c4453c90910a3021546f160bdf0e4d39d1f45bfbc0b"}}, "download_size": 102279, "post_processing_size": null, "dataset_size": 222295, "size_in_bytes": 324574}, "tatoeba.deu": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.deu", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 225583, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.deu-eng.deu": {"num_bytes": 57121, "checksum": "edfa6f75a42554df388f45891735d5e4214158a99def4b73b5908af4a3054551"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.deu-eng.eng": {"num_bytes": 48446, "checksum": "eb9cc83a42f2c4b22f310d05311207e41abb56c2a084666cac3ee0f84d2d0b84"}}, "download_size": 105567, "post_processing_size": null, "dataset_size": 225583, "size_in_bytes": 331150}, "tatoeba.cmn": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.cmn", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 188947, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.cmn-eng.cmn": {"num_bytes": 33410, "checksum": "965d033966fcd186c89741ad49ab4b0a0b2bbd33e02666635ff3b2be23c1ac1f"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.cmn-eng.eng": {"num_bytes": 35521, "checksum": "29ba36232488f5806aceccac57c59c5e750ddd08edb40eef417e3ada9ff9a239"}}, "download_size": 68931, "post_processing_size": null, "dataset_size": 188947, "size_in_bytes": 257878}, "tatoeba.ell": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.ell", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 198977, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.ell-eng.ell": {"num_bytes": 52202, "checksum": "4fb3d3d30bdafd15100dfad5c4680f8f2ed5ca87ed0a6122e2fe2aa21fee65e8"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.ell-eng.eng": {"num_bytes": 26759, "checksum": "1f630710e718e2f85a4e757d3c7f3d6e78ded0b25c99653b4c552138318d9ffe"}}, "download_size": 78961, "post_processing_size": null, "dataset_size": 198977, "size_in_bytes": 277938}, "tatoeba.est": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.est", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 179744, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.est-eng.est": {"num_bytes": 29996, "checksum": "94856999ef35e5357502d7ecf50419d0108b99270e507d9c57f8b283bd1be9c5"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.est-eng.eng": {"num_bytes": 29732, "checksum": "bb033b10596178452aecf2d97ad25580251375e7d224d8b38dad3d93d69b1e4f"}}, "download_size": 59728, "post_processing_size": null, "dataset_size": 179744, "size_in_bytes": 239472}, "tatoeba.eus": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.eus", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 186084, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.eus-eng.eus": {"num_bytes": 34060, "checksum": "4255fb70a6c268b09fcc59a9b308f0fcaaf45ef45e66fc55bf3c80eac4d8c97b"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.eus-eng.eng": {"num_bytes": 32008, "checksum": "1e80f0174ad544697fd69ddcf21287ca10c5e3cacba2fc42bf1d68c460d14ba2"}}, "download_size": 66068, "post_processing_size": null, "dataset_size": 186084, "size_in_bytes": 252152}, "tatoeba.fin": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.fin", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 195685, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.fin-eng.fin": {"num_bytes": 39857, "checksum": "8db3c734f755d578445947f1182f40faf2a9a0eca37561dd248717c088802d60"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.fin-eng.eng": {"num_bytes": 35812, "checksum": "322e610359f4d24852e673bbe4524d52c26dbf980aca0760e95c66dc21ecd504"}}, "download_size": 75669, "post_processing_size": null, "dataset_size": 195685, "size_in_bytes": 271354}, "tatoeba.fra": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.fra", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 200034, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.fra-eng.fra": {"num_bytes": 43727, "checksum": "644172ff9642fefa9e41c29b7d6f44196518e84350dc44d4992e943c0cca92b6"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.fra-eng.eng": {"num_bytes": 36291, "checksum": "5634220f8a26a9a23b84753a9aec0b0832e6bdaa9da3f83e0bd84c928c3f46e3"}}, "download_size": 80018, "post_processing_size": null, "dataset_size": 200034, "size_in_bytes": 280052}, "tatoeba.heb": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.heb", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 203516, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.heb-eng.heb": {"num_bytes": 47660, "checksum": "4a07ca4b8a6fb7ab499791573a2454730f47acbe209359d7b9372a9f6094a102"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.heb-eng.eng": {"num_bytes": 35840, "checksum": "73c27ed0f76c1d2da199230f05489749e10bb67ab879c6dfee8ca9807d6bd99c"}}, "download_size": 83500, "post_processing_size": null, "dataset_size": 203516, "size_in_bytes": 287016}, "tatoeba.hin": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.hin", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 242574, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.hin-eng.hin": {"num_bytes": 88468, "checksum": "15e4fb0a394be4438319f1d6955d1aea226e2a8c5ad38798b23b76ae43d742ed"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.hin-eng.eng": {"num_bytes": 34090, "checksum": "adfce2269a55dbac69b25c0f4f6eb89e0f9383165485925a2e042e61b9480562"}}, "download_size": 122558, "post_processing_size": null, "dataset_size": 242574, "size_in_bytes": 365132}, "tatoeba.hun": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.hun", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 188905, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.hun-eng.hun": {"num_bytes": 35335, "checksum": "56bd0682be8c1db6568313650b3310d641cc8d0019d12dd7caf201302350eeac"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.hun-eng.eng": {"num_bytes": 33554, "checksum": "5323266e91ddee67ed1ae00d6bbac0cdf3d37749d1b2da3459bf0d424bc71383"}}, "download_size": 68889, "post_processing_size": null, "dataset_size": 188905, "size_in_bytes": 257794}, "tatoeba.ind": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.ind", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 194860, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.ind-eng.ind": {"num_bytes": 39969, "checksum": "4f03cd70cba071f746eacd3ebf6b60fd5a8377ce18b4cc52edec6721f05f352a"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.ind-eng.eng": {"num_bytes": 34875, "checksum": "62a3f5127b60eb3526b8fa3994a68fa1a1f114f3a395307a8808a3517c05ffc5"}}, "download_size": 74844, "post_processing_size": null, "dataset_size": 194860, "size_in_bytes": 269704}, "tatoeba.ita": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.ita", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 185849, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.ita-eng.ita": {"num_bytes": 34237, "checksum": "9b3f369d0ed92273b46dd3b983721636e3d15024ce7125f5103229249c386d26"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.ita-eng.eng": {"num_bytes": 31596, "checksum": "738bf8f981e42d285f4f08bc09238782d285b02050e6e95287aa4e998bb7b24b"}}, "download_size": 65833, "post_processing_size": null, "dataset_size": 185849, "size_in_bytes": 251682}, "tatoeba.jav": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.jav", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 38529, "num_examples": 205, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.jav-eng.jav": {"num_bytes": 7457, "checksum": "ad88399db8f94c2a040aa53e7e862225964fac9308a3beb3d5b38f3eca2f827f"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.jav-eng.eng": {"num_bytes": 6456, "checksum": "172776353690f6c047ea21da969fa6979980d692fff1cfbac17eb25851423760"}}, "download_size": 13913, "post_processing_size": null, "dataset_size": 38529, "size_in_bytes": 52442}, "tatoeba.jpn": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.jpn", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 213099, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.jpn-eng.jpn": {"num_bytes": 53844, "checksum": "56040bd6949170a631039d9f8f4c6440db8761b0065c9686feba55c99a320d46"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.jpn-eng.eng": {"num_bytes": 39239, "checksum": "b42129b34e1bf225ccc25fc00e532a6113af98adbc6605b93021bd8aadeb68b6"}}, "download_size": 93083, "post_processing_size": null, "dataset_size": 213099, "size_in_bytes": 306182}, "tatoeba.kat": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.kat", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 161696, "num_examples": 746, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.kat-eng.kat": {"num_bytes": 50967, "checksum": "6ef69b5efbf355597ed91eb355b33a5f524bdf0875dbeaaccf6375badc20e29b"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.kat-eng.eng": {"num_bytes": 21193, "checksum": "d70a14aa64fd7c6b545f11aea754a632e1cbecb91af27fcf6a98a8449a48a8e7"}}, "download_size": 72160, "post_processing_size": null, "dataset_size": 161696, "size_in_bytes": 233856}, "tatoeba.kaz": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.kaz", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 116194, "num_examples": 575, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.kaz-eng.kaz": {"num_bytes": 29687, "checksum": "f20c682582a80b6aa10f3b933db93bc314449b554ce611e263bc75990b319aef"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.kaz-eng.eng": {"num_bytes": 17491, "checksum": "0ab684e7032c6520540d5785adf00ef206d097221d0dd4dc9bcaabd64068e10d"}}, "download_size": 47178, "post_processing_size": null, "dataset_size": 116194, "size_in_bytes": 163372}, "tatoeba.kor": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.kor", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 199155, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.kor-eng.kor": {"num_bytes": 44054, "checksum": "e550c84184ec35b1a0dab3154284719511a21746e53c40f46eb6ab08179e9188"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.kor-eng.eng": {"num_bytes": 35085, "checksum": "f900cf3c9b72ed5a400e1804702863ff3df00be58eb060902e02285d0e68fab3"}}, "download_size": 79139, "post_processing_size": null, "dataset_size": 199155, "size_in_bytes": 278294}, "tatoeba.mal": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.mal", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 177173, "num_examples": 687, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.mal-eng.mal": {"num_bytes": 72952, "checksum": "1a896f54f85a454fb0123864049c65921ae9dfd0cafda6deef8060f0104d965e"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.mal-eng.eng": {"num_bytes": 21765, "checksum": "4d80cdbb844cd4e33f874e5dc45c1cdda4f80998034448f7eb56b8b6532a6622"}}, "download_size": 94717, "post_processing_size": null, "dataset_size": 177173, "size_in_bytes": 271890}, "tatoeba.mar": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.mar", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 220558, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.mar-eng.mar": {"num_bytes": 72652, "checksum": "b2931584fbe62062beb97cc939e4d208ace5ee56f15808860ab14e130fd3c576"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.mar-eng.eng": {"num_bytes": 27890, "checksum": "709d09b697dca053c814b9d525b72cb47cb011aa860c6598f3e2b1b3dd1280dd"}}, "download_size": 100542, "post_processing_size": null, "dataset_size": 220558, "size_in_bytes": 321100}, "tatoeba.nld": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.nld", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 193279, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.nld-eng.nld": {"num_bytes": 37866, "checksum": "d564d4ce1c621ccaefdbe9f5cb08eacccc7bf2a0b58666303e84ca9c7973bdb7"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.nld-eng.eng": {"num_bytes": 35397, "checksum": "3b8836749df573a53235b85ed6771f31bf2de428f520d2d6a1dd94b61b4ef057"}}, "download_size": 73263, "post_processing_size": null, "dataset_size": 193279, "size_in_bytes": 266542}, "tatoeba.pes": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.pes", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 213735, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.pes-eng.pes": {"num_bytes": 58866, "checksum": "f1553713723491fe5876e1060b18fb4abf0c77be3ba06db2e3307e83aedbbb32"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.pes-eng.eng": {"num_bytes": 34853, "checksum": "b5c6cf8c8d93ff8f2fe26b53f3ee29b62db9c6f7dcddcb086ba48953f4ce926b"}}, "download_size": 93719, "post_processing_size": null, "dataset_size": 213735, "size_in_bytes": 307454}, "tatoeba.por": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.por", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 195201, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.por-eng.por": {"num_bytes": 39331, "checksum": "b0c926a232c9889a87a1a970f9399c5618c2d95baf204321e9da794c0aec16f5"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.por-eng.eng": {"num_bytes": 35854, "checksum": "deb4568cfb7b7cbbc060a7fe97c4639fb4680842f4fcd28df791ffdbb753855a"}}, "download_size": 75185, "post_processing_size": null, "dataset_size": 195201, "size_in_bytes": 270386}, "tatoeba.rus": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.rus", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 212488, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.rus-eng.rus": {"num_bytes": 58822, "checksum": "446ff2cae66053c2277d9735b2c2df6b786cae258385f7ade7bed68d8835c6a0"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.rus-eng.eng": {"num_bytes": 33650, "checksum": "7b26f52d6085b7c4944d6f5f6f5b6e1932085b42112f1444db515ce59e878fb8"}}, "download_size": 92472, "post_processing_size": null, "dataset_size": 212488, "size_in_bytes": 304960}, "tatoeba.spa": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.spa", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 192282, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.spa-eng.spa": {"num_bytes": 37490, "checksum": "f9628cea40481e8251f0999718bd893cff0f261752f5e526b3bc20284e2ca018"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.spa-eng.eng": {"num_bytes": 34776, "checksum": "89e4470f4572040b1ca94b3edad97dcd8bd2f0141f072e12933b8659dadf917d"}}, "download_size": 72266, "post_processing_size": null, "dataset_size": 192282, "size_in_bytes": 264548}, "tatoeba.swh": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.swh", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 67283, "num_examples": 390, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.swh-eng.swh": {"num_bytes": 9645, "checksum": "1c672915446c336cc378676e6dbf91eb54d27bbfd0c61563d349265bc6374753"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.swh-eng.eng": {"num_bytes": 10822, "checksum": "e8539647caff9e329776ae863b6224d432923c6e4e9256b9df92ca58ff282eac"}}, "download_size": 20467, "post_processing_size": null, "dataset_size": 67283, "size_in_bytes": 87750}, "tatoeba.tam": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.tam", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 76297, "num_examples": 307, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.tam-eng.tam": {"num_bytes": 30553, "checksum": "bde87fb1ddedccf6c7a2b70ffdd19a959a573c113c2e7a041c4b623fb2170bde"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.tam-eng.eng": {"num_bytes": 8888, "checksum": "e7c5106acdd100214a161970b2a5c31e7386e5b6a963e3d3afdf30412c90ac53"}}, "download_size": 39441, "post_processing_size": null, "dataset_size": 76297, "size_in_bytes": 115738}, "tatoeba.tel": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.tel", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 53239, "num_examples": 234, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.tel-eng.tel": {"num_bytes": 18337, "checksum": "7e1a1bcd106cce650a09e2f042f1354b55b29bea2bcfa86554dfa0ad12ce8976"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.tel-eng.eng": {"num_bytes": 6806, "checksum": "1efc2ef57d9b1ecebfc4baa45e86fd793e38473304e9c043aebabc3a1b29a294"}}, "download_size": 25143, "post_processing_size": null, "dataset_size": 53239, "size_in_bytes": 78382}, "tatoeba.tgl": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.tgl", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 188154, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.tgl-eng.tgl": {"num_bytes": 36506, "checksum": "f99165dc05190b99f6574fe24db884ff85d111612a25e7a37323f001aafc2a6e"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.tgl-eng.eng": {"num_bytes": 31632, "checksum": "e7c9beda3f3072a968a34a7226a66d1ebf1dcb33cf002805dc752f80a7c620ae"}}, "download_size": 68138, "post_processing_size": null, "dataset_size": 188154, "size_in_bytes": 256292}, "tatoeba.tha": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.tha", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 128974, "num_examples": 548, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.tha-eng.tha": {"num_bytes": 44759, "checksum": "65c7b3c01a56a1ac8971e72e0ea8e74a027718dc84044d8802c0ab36395a3156"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.tha-eng.eng": {"num_bytes": 18439, "checksum": "2881c82d2c5fa59cf0a68bc9e012f5e0b0a716f7357cbecf77c247efc2fd7294"}}, "download_size": 63198, "post_processing_size": null, "dataset_size": 128974, "size_in_bytes": 192172}, "tatoeba.tur": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.tur", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 191901, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.tur-eng.tur": {"num_bytes": 37607, "checksum": "1ffa0acc006018b3105abda41a4d4ca42f3c122964a49b71793546367b079a86"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.tur-eng.eng": {"num_bytes": 34278, "checksum": "a768df7dd3d1344f872a458b32c3a65e24f8381826ccb16ba6677426176c8121"}}, "download_size": 71885, "post_processing_size": null, "dataset_size": 191901, "size_in_bytes": 263786}, "tatoeba.urd": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.urd", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 208728, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.urd-eng.urd": {"num_bytes": 56819, "checksum": "2efc22dc61885a9454aeeee68c8b841c7f9138d53ba644a82308bd210140450b"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.urd-eng.eng": {"num_bytes": 31893, "checksum": "dab35fda3f73b3fd86b6b9f9f9f6242430961aa5d1ac247adbc646867df79cec"}}, "download_size": 88712, "post_processing_size": null, "dataset_size": 208728, "size_in_bytes": 297440}, "tatoeba.vie": {"description": "his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.\nFor each languages, we have selected 1000 English sentences and their translations, if available. Please check\nthis paper for a description of the languages, their families and scripts as well as baseline results.\nPlease note that the English sentences are not identical for all language pairs. This means that the results are\nnot directly comparable across languages. In particular, the sentences tend to have less variety for several\nlow-resource languages, e.g. \"Tom needed water\", \"Tom needs water\", \"Tom is getting water\", ...\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": " @article{tatoeba,\n title={Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},\n author={Mikel, Artetxe and Holger, Schwenk,},\n journal={arXiv:1812.10464v2},\n year={2018}\n}\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md", "license": "", "features": {"source_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "target_sentence": {"dtype": "string", "id": null, "_type": "Value"}, "source_lang": {"dtype": "string", "id": null, "_type": "Value"}, "target_lang": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "tatoeba.vie", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 211423, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.vie-eng.vie": {"num_bytes": 52721, "checksum": "6dbb02d778b0bfc8678cd85f87db76de55dd7e409a26fe32ad42d50e0f1fff77"}, "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/tatoeba.vie-eng.eng": {"num_bytes": 38686, "checksum": "a1f60bd8ae6c42224a4c050d2aa1ff4242d14827d64d7831e96ecf2b2c367f5f"}}, "download_size": 91407, "post_processing_size": null, "dataset_size": 211423, "size_in_bytes": 302830}, "udpos.Afrikaans": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Afrikaans", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 91302, "num_examples": 194, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 174256, "num_examples": 425, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 586382, "num_examples": 1315, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 851940, "size_in_bytes": 356068621}, "udpos.Arabic": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Arabic", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 593662, "num_examples": 909, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 973834, "num_examples": 1680, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 4453694, "num_examples": 6075, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 6021190, "size_in_bytes": 361237871}, "udpos.Basque": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Basque", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 438683, "num_examples": 1798, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 444656, "num_examples": 1799, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 1327725, "num_examples": 5396, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 2211064, "size_in_bytes": 357427745}, "udpos.Bulgarian": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Bulgarian", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 347129, "num_examples": 1115, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 339959, "num_examples": 1116, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 2689779, "num_examples": 8907, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 3376867, "size_in_bytes": 358593548}, "udpos.Dutch": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Dutch", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 393604, "num_examples": 1394, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 397916, "num_examples": 1471, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 4518018, "num_examples": 18051, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 5309538, "size_in_bytes": 360526219}, "udpos.English": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.English", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 1042052, "num_examples": 3974, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 1421160, "num_examples": 5440, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 6225545, "num_examples": 21253, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 8688757, "size_in_bytes": 363905438}, "udpos.Estonian": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Estonian", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 814183, "num_examples": 3125, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 1065713, "num_examples": 3760, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 6614929, "num_examples": 25749, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 8494825, "size_in_bytes": 363711506}, "udpos.Finnish": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Finnish", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 656658, "num_examples": 3239, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 1025738, "num_examples": 4422, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 5613742, "num_examples": 27198, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 7296138, "size_in_bytes": 362512819}, "udpos.French": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.French", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 1294108, "num_examples": 5979, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 1731061, "num_examples": 9465, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 10118993, "num_examples": 47308, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 13144162, "size_in_bytes": 368360843}, "udpos.German": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.German", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 6044862, "num_examples": 19233, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 7345899, "num_examples": 22458, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 54773981, "num_examples": 166849, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 68164742, "size_in_bytes": 423381423}, "udpos.Greek": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Greek", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 1062459, "num_examples": 2559, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 1028677, "num_examples": 2809, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 8932140, "num_examples": 28152, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 11023276, "size_in_bytes": 366239957}, "udpos.Hebrew": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Hebrew", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 210025, "num_examples": 484, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 223877, "num_examples": 491, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 2505703, "num_examples": 5241, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 2939605, "size_in_bytes": 358156286}, "udpos.Hindi": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Hindi", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 839714, "num_examples": 1659, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 1400237, "num_examples": 2684, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 6690274, "num_examples": 13304, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 8930225, "size_in_bytes": 364146906}, "udpos.Hungarian": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Hungarian", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 215891, "num_examples": 441, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 193740, "num_examples": 449, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 372238, "num_examples": 910, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 781869, "size_in_bytes": 355998550}, "udpos.Indonesian": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Indonesian", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 220875, "num_examples": 559, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 557113, "num_examples": 1557, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 1710690, "num_examples": 4477, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 2488678, "size_in_bytes": 357705359}, "udpos.Italian": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Italian", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 989008, "num_examples": 2278, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 1337881, "num_examples": 3518, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 11299329, "num_examples": 29685, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 13626218, "size_in_bytes": 368842899}, "udpos.Japanese": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Japanese", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 200368, "num_examples": 511, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 928914, "num_examples": 2372, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 2792963, "num_examples": 7125, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 3922245, "size_in_bytes": 359138926}, "udpos.Kazakh": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Kazakh", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 228936, "num_examples": 1047, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 11450, "num_examples": 31, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 240386, "size_in_bytes": 355457067}, "udpos.Korean": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Korean", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 782599, "num_examples": 3016, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 1162551, "num_examples": 4276, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 7341303, "num_examples": 27410, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 9286453, "size_in_bytes": 364503134}, "udpos.Chinese": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Chinese", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 594460, "num_examples": 3038, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 1236063, "num_examples": 5528, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 4218915, "num_examples": 18998, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 6049438, "size_in_bytes": 361266119}, "udpos.Marathi": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Marathi", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 8509, "num_examples": 46, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 7883, "num_examples": 47, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 59035, "num_examples": 373, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 75427, "size_in_bytes": 355292108}, "udpos.Persian": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Persian", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 317065, "num_examples": 599, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 320695, "num_examples": 600, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 2400788, "num_examples": 4798, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 3038548, "size_in_bytes": 358255229}, "udpos.Portuguese": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Portuguese", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 712409, "num_examples": 1770, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 1082594, "num_examples": 2681, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 7669580, "num_examples": 17992, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 9464583, "size_in_bytes": 364681264}, "udpos.Russian": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Russian", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 3457043, "num_examples": 9960, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 4236717, "num_examples": 11336, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 24230182, "num_examples": 67435, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 31923942, "size_in_bytes": 387140623}, "udpos.Spanish": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Spanish", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 1498777, "num_examples": 3054, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 1476512, "num_examples": 3147, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 13858442, "num_examples": 28492, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 16833731, "size_in_bytes": 372050412}, "udpos.Tagalog": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Tagalog", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 5165, "num_examples": 55, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 5165, "size_in_bytes": 355221846}, "udpos.Tamil": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Tamil", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 40043, "num_examples": 80, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 62378, "num_examples": 120, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 202608, "num_examples": 400, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 305029, "size_in_bytes": 355521710}, "udpos.Telugu": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Telugu", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 18002, "num_examples": 131, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 19587, "num_examples": 146, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 138061, "num_examples": 1051, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 175650, "size_in_bytes": 355392331}, "udpos.Thai": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Thai", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 561348, "num_examples": 1000, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 561348, "size_in_bytes": 355778029}, "udpos.Turkish": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Turkish", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 186467, "num_examples": 988, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 827394, "num_examples": 4785, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 704417, "num_examples": 3664, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 1718278, "size_in_bytes": 356934959}, "udpos.Urdu": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Urdu", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 284273, "num_examples": 552, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 288565, "num_examples": 535, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 2107374, "num_examples": 4043, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 2680212, "size_in_bytes": 357896893}, "udpos.Vietnamese": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Vietnamese", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 206200, "num_examples": 800, "dataset_name": "xtreme"}, "test": {"name": "test", "num_bytes": 214075, "num_examples": 800, "dataset_name": "xtreme"}, "train": {"name": "train", "num_bytes": 367347, "num_examples": 1400, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 787622, "size_in_bytes": 356004303}, "udpos.Yoruba": {"description": "Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological\nfeatures, and syntactic dependencies) across different human languages. UD is an open community effort with over 200\ncontributors producing more than 100 treebanks in over 70 languages. If you\u2019re new to UD, you should start by reading\nthe first part of the Short Introduction and then browsing the annotation guidelines.\n\nThe Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of\nthe cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages\n(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of\nsyntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,\nand availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil\n(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the\nNiger-Congo languages Swahili and Yoruba, spoken in Africa.\n", "citation": "\n@article{hu2020xtreme,\n author = {Junjie Hu and Sebastian Ruder and Aditya Siddhant and Graham Neubig and Orhan Firat and Melvin Johnson},\n title = {XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization},\n journal = {CoRR},\n volume = {abs/2003.11080},\n year = {2020},\n archivePrefix = {arXiv},\n eprint = {2003.11080}\n}\n", "homepage": "https://github.com/google-research/xtreme\thttps://universaldependencies.org/", "license": "", "features": {"tokens": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "pos_tags": {"feature": {"num_classes": 17, "names": ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"], "id": null, "_type": "ClassLabel"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "xtreme", "config_name": "udpos.Yoruba", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"test": {"name": "test", "num_bytes": 44668, "num_examples": 100, "dataset_name": "xtreme"}}, "download_checksums": {"https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz": {"num_bytes": 355216681, "checksum": "5ff973e44345a5f69b94cc1427158e14e851c967d58773cc2ac5a1d3adaca409"}}, "download_size": 355216681, "post_processing_size": null, "dataset_size": 44668, "size_in_bytes": 355261349}} \ No newline at end of file