oscar-corpus/OSCAR-2201

Dataset Card for "oscar"

Dataset Summary

OSCAR or Open Super-large Crawled Aggregated coRpus is a huge multilingual corpus obtained by language classification and filtering of the Common Crawl corpus using the ungoliant architecture. Data is distributed by language in both original and deduplicated form.

We are aware of the virus warnings issue. See discussion here for more info!

Usage

from datasets import load_dataset


dataset = load_dataset("oscar-corpus/OSCAR-2201",
                        use_auth_token=True, # required
                        language="ar", 
                        streaming=True, # optional
                        split="train") # optional, but the dataset only has a train split

for d in dataset:
    print(d) # prints documents

Supported Tasks and Leaderboards

OSCAR is mainly intended to pretrain language models and word representations.

Languages

All the data is distributed by language, both the original and the deduplicated versions of the data are available. 151 different languages are available. The table in subsection Data Splits Sample Size provides the language code for each subcorpus as well as the number of words (space separated tokens), lines and sizes for both the original and the deduplicated versions of OSCAR.

Issues

OSCAR 22.01 may have quality issues on low size subcorpora, as it has been the case before.

Note that since the documents are identified as a whole, it is expected to have lines in other languages in a given language subcorpus. As an example, it is known and expected that the German subcorpus contains documents holding lines identified as Swiss German / Alemannic.

If you encounter something that is unexpected, please file an issue here: https://github.com/oscar-corpus/corpus/issues.

Language code	Language	Issues

Dataset Structure

We show detailed information for all the configurations of the dataset.

Data Instances

TODO

Data Fields

• id: a int64 feature.

• content: string Newline-separated content

• warc_headers: WARC Headers

• warc_headers.content-length: int64 Content length (in bytes) before cleaning

• warc_headers.content-type: string MIME type

• warc_headers.warc-block-digest:string Algorithm name and calculated value of a digest applied to the full block of the record

• warc_headers.warc-date: string Crawl date (YYYY-MM-DDThh:mm:ssZ)

• warc_headers.warc-identified-content-language: string Comma-separated list of language identifications done by CommonCrawl (uses CLD3)

• warc_headers.warc-record-id: string Record ID

• warc_headers.warc-refers-to: string Record-ID of a single record for which the present record holds additional content

• warc_headers.warc-target-uri: string URI from where the content has been fetched

• warc_headers.warc-type: string Type of the WARC Record

• metadata: Metadata

• metadata.identification.label: string Language identification of the document

• metadata.identification.prob: float Confidence of the identification

• metadata.annotation: [string] Annnotations of the document. null if none present. (Is None if using datasets)

• metadata.sentence_identifications: [string] List of line identifications. null/None can be present for lines that failed the identification step.

• meta.offset: int64 line offset where the related text begins. Should be used with meta.nb_sentences when reading the source files rather than using iterators to get related data.

• text: string content

See the WARC Format standard for more details on the warc_headers fields, and our website for more details about the format in general.

Data Splits

Click to expand the number of samples per configuration

Table

lang	size	docs	words
Multilingual	12.1 GB	1,210,685	936,187,711
Afrikaans	47.0 MB	12,393	6,227,310
Albanian	3.0 GB	437,287	326,325,149
Alemannic / Swiss German	363.6 kB	139	37,381
Amharic	461.0 MB	37,513	30,481,153
Arabic	84.2 GB	8,718,929	6,103,711,887
Aragonese	10.6 kB	12	51
Armenian	4.7 GB	379,267	268,031,270
Assamese	221.2 MB	17,084	11,109,557
Asturian	73.6 kB	77	3,919
Avaric	18.6 kB	14	582
Azerbaijani	3.5 GB	491,847	291,927,692
Bangla	15.1 GB	1,171,501	751,877,226
Bashkir	95.5 MB	11,198	5,418,474
Basque	1.1 GB	233,658	97,092,942
Belarusian	1.8 GB	180,046	107,227,860
Bihari languages	24.2 kB	27	569
Bishnupriya	2.0 MB	271	98,419
Bosnian	10.3 kB	10	422
Breton	33.7 MB	16,119	3,111,619
Bulgarian	35.1 GB	2,887,115	2,405,981,285
Burmese	1.9 GB	158,733	44,835,970
Catalan	13.9 GB	2,627,307	1,508,919,864
Cebuano	44.6 MB	5,742	5,253,785
Central Kurdish	716.4 MB	84,950	43,913,025
Chechen	14.0 MB	4,086	798,766
Chinese	900.9 GB	56,524,518	23,149,203,886
Chuvash	41.8 MB	4,750	2,465,782
Cornish	1.4 kB	2	55
Croatian	11.2 MB	11,462	505,369
Czech	58.6 GB	10,381,916	5,452,724,456
Danish	12.6 GB	2,265,479	1,454,439,292
Dimli (individual language)	706 Bytes	1	19
Divehi	217.2 MB	24,067	10,112,205
Dutch	114.0 GB	20,206,532	12,329,127,151
Eastern Mari	11.3 MB	1,612	641,525
Egyptian Arabic	2.8 MB	1,256	176,096
English	3.2 TB	431,992,659	377,376,402,775
Esperanto	558.3 MB	111,932	58,416,628
Estonian	9.2 GB	1,362,524	820,975,443
Filipino	646.5 MB	70,394	81,881,278
Finnish	37.8 GB	4,948,961	2,900,615,928
French	382.2 GB	52,037,098	41,713,990,658
Galician	255.2 MB	88,803	27,051,212
Georgian	7.1 GB	488,588	281,430,479
German	496.7 GB	70,075,424	46,826,676,844
Goan Konkani	787.2 kB	46	38,831
Greek	78.3 GB	6,738,546	5,031,242,803
Guarani	9.0 kB	10	374
Gujarati	4.8 GB	136,467	301,170,777
Hebrew	30.3 GB	3,132,396	2,249,377,984
Hindi	23.3 GB	1,529,907	1,534,799,198
Hungarian	53.9 GB	6,866,062	4,598,787,907
Icelandic	2.0 GB	396,183	210,365,124
Ido	77.3 kB	105	2,690
Iloko	97.9 kB	75	8,592
Indonesian	17.4 GB	2,244,622	1,984,195,207
Interlingua	40.2 kB	6	10,125
Irish	45.6 MB	12,233	4,877,850
Italian	229.3 GB	28,502,092	24,294,684,830
Japanese	258.7 GB	36,328,931	5,592,948,356
Javanese	152.7 kB	70	10,441
Kalmyk	9.3 kB	9	250
Kannada	2.6 GB	150,850	108,450,571
Karachay-Balkar	119.6 kB	91	4,089
Kazakh	2.9 GB	261,085	157,267,307
Khmer	1.9 GB	121,910	30,564,131
Komi	119.9 kB	127	3,335
Korean	51.8 GB	5,881,481	3,854,968,649
Kurdish	150.3 MB	29,906	17,390,759
Kyrgyz	518.6 MB	62,244	28,028,986
Lao	337.1 MB	28,914	6,682,982
Latin	4.1 MB	4,397	187,446
Latvian	8.2 GB	1,032,987	707,361,898
Lezghian	375.5 kB	124	19,250
Limburgish	1.4 kB	2	41
Lithuanian	20.0 GB	2,303,070	1,712,802,056
Lojban	1.9 MB	570	260,542
Lombard	2.6 kB	2	225
Low German	9.0 MB	1,938	1,012,561
Lower Sorbian	707 Bytes	1	17
Luxembourgish	15.8 MB	5,108	1,545,946
Macedonian	3.6 GB	341,775	244,058,579
Maithili	21.6 kB	23	483
Malagasy	57.3 MB	3,028	7,279,056
Malay	5.3 MB	5,228	217,818
Malayalam	4.1 GB	250,972	137,831,247
Maltese	2.5 MB	2,208	118,190
Marathi	3.3 GB	250,376	160,179,233
Mazanderani	128.2 kB	76	7,337
Minangkabau	6.0 MB	585	614,613
Mingrelian	7.6 MB	2,550	253,333
Mongolian	2.8 GB	237,719	176,405,432
Nahuatl languages	8.7 kB	12	179
Nepali	3.7 GB	391,947	177,885,116
Newari	5.7 MB	1,134	273,837
Norwegian	2.8 GB	973,188	279,182,902
Norwegian Nynorsk	6.8 MB	5,835	459,183
Occitan	2.1 MB	373	31,061
Odia	487.9 MB	52,942	23,755,902
Ossetic	13.9 MB	3,560	800,430
Pashto	490.3 MB	50,312	46,293,249
Persian	77.4 GB	7,665,871	6,430,164,396
Piedmontese	1.7 MB	698	188,270
Polish	139.0 GB	19,301,137	12,584,498,906
Portuguese	170.3 GB	23,735,707	18,441,864,893
Punjabi	1.1 GB	68,094	70,068,604
Quechua	744 Bytes	1	14
Romanian	49.2 GB	4,624,764	5,261,803,995
Russia Buriat	32.9 kB	39	785
Russian	1.1 TB	76,060,844	62,811,122,663
Sakha	65.6 MB	6,284	3,473,813
Sanskrit	136.0 MB	4,472	5,671,369
Scottish Gaelic	137.7 kB	136	7,769
Serbian	6.9 GB	577,472	482,932,670
Serbian (Latin)	931.8 kB	738	92,875
Sicilian	1.5 kB	2	50
Sindhi	117.1 MB	15,516	10,685,611
Sinhala	2.0 GB	108,593	113,179,741
Slovak	16.5 GB	2,409,555	1,619,121,944
Slovenian	1.2 GB	351,894	118,400,246
Somali	2.1 kB	3	109
South Azerbaijani	14.1 MB	5,381	693,746
Spanish	381.9 GB	51,386,247	42,829,835,316
Sundanese	5.0 MB	263	547,145
Swahili	1.3 MB	462	123,050
Swedish	48.0 GB	7,541,278	5,078,331,128
Tajik	870.9 MB	46,366	56,627,727
Tamil	11.4 GB	556,772	452,343,748
Tatar	915.3 MB	76,398	51,875,265
Telugu	3.4 GB	249,756	137,752,065
Thai	66.1 GB	5,030,254	1,626,779,846
Tibetan	234.5 MB	18,683	2,286,269
Turkish	75.1 GB	10,826,031	6,421,221,358
Turkmen	4.4 MB	2,485	276,632
Ukrainian	48.8 GB	4,558,214	2,879,585,992
Emiliano-Romagnolo[eml]	901 Bytes	1	53
Upper Sorbian	132.8 kB	110	8,825
Urdu	3.4 GB	336,994	332,816,354
Uyghur	201.9 MB	18,556	11,240,889
Uzbek	19.9 MB	9,526	1,370,842
Vietnamese	98.9 GB	9,587,233	12,283,185,482
Volapük	825.9 kB	661	57,039
Walloon	105.7 kB	138	4,386
Waray	7.6 MB	933	830,872
Welsh	409.3 MB	90,378	49,488,495
Western Frisian	75.3 MB	21,946	6,357,929
Western Mari	743.5 kB	155	43,916
Western Panjabi	46.7 MB	6,790	4,060,419
Wu Chinese	137.2 kB	88	3,056
Yiddish	232.5 MB	23,418	15,809,780
Yoruba	24.7 kB	26	1,042

Dataset Creation

Curation Rationale

OSCAR was constructed using Ungoliant, a new pipeline derived from goclassy, itself being derived from fastText's one.

The pipeline works on documents rather than lines. Ungoliant is implemented in the Rust programming language, and uses rayon as its data parallelism strategy. Threading is done at shard, record and sentence level, making the whole generation process much more efficient.

Filtering will be explained in a future blog post at our website

Source Data

Initial Data Collection and Normalization

Common Crawl is a non-profit foundation which produces and maintains an open repository of web crawled data that is both accessible and analysable. Common Crawl's complete web archive consists of petabytes of data collected over 8 years of web crawling. The repository contains raw web page HTML data (WARC files), metdata extracts (WAT files) and plain text extracts (WET files). The organisation's crawlers has always respected nofollow and robots.txt policies.

Each monthly Common Crawl snapshot is in itself a massive multilingual corpus, where every single file contains data coming from multiple web pages written in a large variety of languages and covering all possible types of topics.

To construct OSCAR the WET files of Common Crawl were used. These contain the extracted plain texts from the websites mostly converted to UTF-8, as well as headers containing the metatada of each crawled document. Each WET file comes compressed in gzip format and is stored on Amazon Web Services. In the case of OSCAR 22.01, the November/December 2021 snapshot was used. It is composed by 64 000 compressed text files containing documents and their headers.

Who are the source language producers?

The data comes from multiple web pages in a large variety of languages.

Annotations

The dataset does not contain any additional annotations.

Annotation process

N/A

Who are the annotators?

N/A

Personal and Sensitive Information

Being constructed from Common Crawl, Personal and sensitive information might be present. This must be considered before training deep learning models with OSCAR, specially in the case of text-generation models.

Considerations for Using the Data

Social Impact of Dataset

OSCAR is intended to bring more data to a wide variety of lanuages, the aim of the corpus is to make large amounts of data available to lower resource languages in order to facilitate the pre-training of state-of-the-art language modeling architectures.

Discussion of Biases

OSCAR is not properly filtered yet and this can be reflected on the models trained with it. Care is advised specially concerning biases of the resulting models.

Other Known Limitations

The fastText linear classifier is limed both in performance and the variety of languages it can recognize, so the quality of some OSCAR sub-corpora might be lower than expected, specially for the lowest-resource langiuages. Some audits have already been done by third parties.

Additional Information

Dataset Curators

The corpus was put together by Julien Abadji, Pedro Ortiz Suarez, Benoît Sagot, and Laurent Romary, during work done at Inria, particularly at the ALMAnaCH team.

Licensing Information

These data are released under this licensing scheme
We do not own any of the text from which these data has been extracted.
We license the actual packaging of these data under the Creative Commons CC0 license ("no rights reserved") http://creativecommons.org/publicdomain/zero/1.0/
To the extent possible under law, Inria has waived all copyright and related or neighboring rights to OSCAR
This work is published from: France.

Should you consider that our data contains material that is owned by you and should therefore not be reproduced here, please:
* Clearly identify yourself, with detailed contact data such as an address, telephone number or email address at which you can be contacted.
* Clearly identify the copyrighted work claimed to be infringed.
* Clearly identify the material that is claimed to be infringing and information reasonably sufficient to allow us to locate the material.

We will comply to legitimate requests by removing the affected sources from the next release of the corpus.

Citation Information

@ARTICLE{2022arXiv220106642A,
       author = {{Abadji}, Julien and {Ortiz Suarez}, Pedro and {Romary}, Laurent and {Sagot}, Beno{\^\i}t},
        title = "{Towards a Cleaner Document-Oriented Multilingual Crawled Corpus}",
      journal = {arXiv e-prints},
     keywords = {Computer Science - Computation and Language},
         year = 2022,
        month = jan,
          eid = {arXiv:2201.06642},
        pages = {arXiv:2201.06642},
archivePrefix = {arXiv},
       eprint = {2201.06642},
 primaryClass = {cs.CL},
       adsurl = {https://ui.adsabs.harvard.edu/abs/2022arXiv220106642A},
      adsnote = {Provided by the SAO/NASA Astrophysics Data System}
}
      
@inproceedings{AbadjiOrtizSuarezRomaryetal.2021,
  author    = {Julien Abadji and Pedro Javier Ortiz Su{\'a}rez and Laurent Romary and Beno{\^i}t Sagot},
  title     = {Ungoliant: An optimized pipeline for the generation of a very large-scale multilingual web corpus},
  series = {Proceedings of the Workshop on Challenges in the Management of Large Corpora (CMLC-9) 2021. Limerick, 12 July 2021 (Online-Event)},
  editor    = {Harald L{\"u}ngen and Marc Kupietz and Piotr Bański and Adrien Barbaresi and Simon Clematide and Ines Pisetta},
  publisher = {Leibniz-Institut f{\"u}r Deutsche Sprache},
  address   = {Mannheim},
  doi       = {10.14618/ids-pub-10468},
  url       = {https://nbn-resolving.org/urn:nbn:de:bsz:mh39-104688},
  pages     = {1 -- 9},
  year      = {2021},
  abstract  = {Since the introduction of large language models in Natural Language Processing, large raw corpora have played a crucial role in Computational Linguistics. However, most of these large raw corpora are either available only for English or not available to the general public due to copyright issues. Nevertheless, there are some examples of freely available multilingual corpora for training Deep Learning NLP models, such as the OSCAR and Paracrawl corpora. However, they have quality issues, especially for low-resource languages. Moreover, recreating or updating these corpora is very complex. In this work, we try to reproduce and improve the goclassy pipeline used to create the OSCAR corpus. We propose a new pipeline that is faster, modular, parameterizable, and well documented. We use it to create a corpus similar to OSCAR but larger and based on recent data. Also, unlike OSCAR, the metadata information is at the document level. We release our pipeline under an open source license and publish the corpus under a research-only license.},
  language  = {en}
}

@ARTICLE{caswell-etal-2021-quality,
       author = {{Caswell}, Isaac and {Kreutzer}, Julia and {Wang}, Lisa and {Wahab}, Ahsan and {van Esch}, Daan and {Ulzii-Orshikh}, Nasanbayar and {Tapo}, Allahsera and {Subramani}, Nishant and {Sokolov}, Artem and {Sikasote}, Claytone and {Setyawan}, Monang and {Sarin}, Supheakmungkol and {Samb}, Sokhar and {Sagot}, Beno{\^\i}t and {Rivera}, Clara and {Rios}, Annette and {Papadimitriou}, Isabel and {Osei}, Salomey and {Ortiz Su{\'a}rez}, Pedro Javier and {Orife}, Iroro and {Ogueji}, Kelechi and {Niyongabo}, Rubungo Andre and {Nguyen}, Toan Q. and {M{\"u}ller}, Mathias and {M{\"u}ller}, Andr{\'e} and {Hassan Muhammad}, Shamsuddeen and {Muhammad}, Nanda and {Mnyakeni}, Ayanda and {Mirzakhalov}, Jamshidbek and {Matangira}, Tapiwanashe and {Leong}, Colin and {Lawson}, Nze and {Kudugunta}, Sneha and {Jernite}, Yacine and {Jenny}, Mathias and {Firat}, Orhan and {Dossou}, Bonaventure F.~P. and {Dlamini}, Sakhile and {de Silva}, Nisansa and {{\c{C}}abuk Ball{\i}}, Sakine and {Biderman}, Stella and {Battisti}, Alessia and {Baruwa}, Ahmed and {Bapna}, Ankur and {Baljekar}, Pallavi and {Abebe Azime}, Israel and {Awokoya}, Ayodele and {Ataman}, Duygu and {Ahia}, Orevaoghene and {Ahia}, Oghenefego and {Agrawal}, Sweta and {Adeyemi}, Mofetoluwa},
        title = "{Quality at a Glance: An Audit of Web-Crawled Multilingual Datasets}",
      journal = {arXiv e-prints},
     keywords = {Computer Science - Computation and Language, Computer Science - Artificial Intelligence},
         year = 2021,
        month = mar,
          eid = {arXiv:2103.12028},
        pages = {arXiv:2103.12028},
archivePrefix = {arXiv},
       eprint = {2103.12028},
 primaryClass = {cs.CL},
       adsurl = {https://ui.adsabs.harvard.edu/abs/2021arXiv210312028C},
      adsnote = {Provided by the SAO/NASA Astrophysics Data System}
}

@inproceedings{ortiz-suarez-etal-2020-monolingual,
    title = "A Monolingual Approach to Contextualized Word Embeddings for Mid-Resource Languages",
    author = "Ortiz Su{'a}rez, Pedro Javier  and
      Romary, Laurent  and
      Sagot, Benoit",
    booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
    month = jul,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.acl-main.156",
    pages = "1703--1714",
    abstract = "We use the multilingual OSCAR corpus, extracted from Common Crawl via language classification, filtering and cleaning, to train monolingual contextualized word embeddings (ELMo) for five mid-resource languages. We then compare the performance of OSCAR-based and Wikipedia-based ELMo embeddings for these languages on the part-of-speech tagging and parsing tasks. We show that, despite the noise in the Common-Crawl-based OSCAR data, embeddings trained on OSCAR perform much better than monolingual embeddings trained on Wikipedia. They actually equal or improve the current state of the art in tagging and parsing for all five languages. In particular, they also improve over multilingual Wikipedia-based contextual embeddings (multilingual BERT), which almost always constitutes the previous state of the art, thereby showing that the benefit of a larger, more diverse corpus surpasses the cross-lingual benefit of multilingual embedding architectures.",
}

@inproceedings{OrtizSuarezSagotRomary2019,
  author    = {Pedro Javier {Ortiz Su{'a}rez} and Benoit Sagot and Laurent Romary},
  title     = {Asynchronous pipelines for processing huge corpora on medium to low resource infrastructures},
  series = {Proceedings of the Workshop on Challenges in the Management of Large Corpora (CMLC-7) 2019. Cardiff, 22nd July 2019},
  editor    = {Piotr Bański and Adrien Barbaresi and Hanno Biber and Evelyn Breiteneder and Simon Clematide and Marc Kupietz and Harald L{"u}ngen and Caroline Iliadi},
  publisher = {Leibniz-Institut f{"u}r Deutsche Sprache},
  address   = {Mannheim},
  doi       = {10.14618/ids-pub-9021},
  url       = {http://nbn-resolving.de/urn:nbn:de:bsz:mh39-90215},
  pages     = {9 -- 16},
  year      = {2019},
  abstract  = {Common Crawl is a considerably large, heterogeneous multilingual corpus comprised of crawled documents from the internet, surpassing 20TB of data and distributed as a set of more than 50 thousand plain text files where each contains many documents written in a wide variety of languages. Even though each document has a metadata block associated to it, this data lacks any information about the language in which each document is written, making it extremely difficult to use Common Crawl for monolingual applications. We propose a general, highly parallel, multithreaded pipeline to clean and classify Common Crawl by language; we specifically design it so that it runs efficiently on medium to low resource infrastructures where I/O speeds are the main constraint. We develop the pipeline so that it can be easily reapplied to any kind of heterogeneous corpus and so that it can be parameterised to a wide range of infrastructures. We also distribute a 6.3TB version of Common Crawl, filtered, classified by language, shuffled at line level in order to avoid copyright issues, and ready to be used for NLP applications.},
  language  = {en}
}

Contributions

Thanks to @pjox, @Uinelj and @lhoestq for adding this dataset.