--- language: - af - ar - bg - ca - cs - da - de - el - en - es - et - eu - fa - fi - fr - ga - he - hi - hr - hu - hy - id - it - ja - ko - lt - lv - mr - nl - 'no' - pl - pt - ro - ru - sa - sk - sl - sr - sv - ta - te - tr - uk - ur - vi license: mit pretty_name: Multilingual Tokenizer Wikipedia Benchmark dataset_info: - config_name: af features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 541481060 num_examples: 112518 - name: clean num_bytes: 539551289.6071739 num_examples: 112117 download_size: 441191361 dataset_size: 1081032349.607174 - config_name: ar features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 7007645793 num_examples: 1219201 - name: clean num_bytes: 6980694657.688122 num_examples: 1214512 download_size: 4415559180 dataset_size: 13988340450.688122 - config_name: bg features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 2437923560 num_examples: 294275 - name: clean num_bytes: 2433855866.6248918 num_examples: 293784 download_size: 1805069655 dataset_size: 4871779426.624891 - config_name: ca features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 4801022979 num_examples: 737409 - name: clean num_bytes: 4766991732.959834 num_examples: 732182 download_size: 3884482903 dataset_size: 9568014711.959835 - config_name: cs features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 3740905267 num_examples: 534044 - name: clean num_bytes: 3730243864.91258 num_examples: 532522 download_size: 3671037924 dataset_size: 7471149131.9125805 - config_name: da features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 1298277678 num_examples: 295347 - name: clean num_bytes: 1292602738.074089 num_examples: 294056 download_size: 1782396281 dataset_size: 2590880416.074089 - config_name: de features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 23086869184 num_examples: 2845308 - name: clean num_bytes: 23073148386.18474 num_examples: 2843617 download_size: 21942020975 dataset_size: 46160017570.18474 - config_name: el features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 3002968703 num_examples: 226834 - name: clean num_bytes: 2973684879.714972 num_examples: 224622 download_size: 2295250961 dataset_size: 5976653582.714972 - config_name: en features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 49746869820 num_examples: 6407814 - name: clean num_bytes: 49560903666.851944 num_examples: 6383860 download_size: 40592018321 dataset_size: 99307773486.85194 - config_name: es features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 14759846818 num_examples: 1841155 - name: clean num_bytes: 14536992695.618353 num_examples: 1813356 download_size: 12175892555 dataset_size: 29296839513.618355 - config_name: et features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 1019050491 num_examples: 240397 - name: clean num_bytes: 1016723262.6254404 num_examples: 239848 download_size: 1019164563 dataset_size: 2035773753.6254404 - config_name: eu features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 1291195010 num_examples: 416347 - name: clean num_bytes: 1265327506.262949 num_examples: 408006 download_size: 968840915 dataset_size: 2556522516.262949 - config_name: fa features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 4224898253 num_examples: 979869 - name: clean num_bytes: 4213433450.6083264 num_examples: 977210 download_size: 2499698548 dataset_size: 8438331703.608326 - config_name: fi features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 2600737260 num_examples: 561598 - name: clean num_bytes: 2595874753.1481237 num_examples: 560548 download_size: 2642007766 dataset_size: 5196612013.148124 - config_name: fr features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 20069732840 num_examples: 2564646 - name: clean num_bytes: 19942544382.860683 num_examples: 2548393 download_size: 16151551755 dataset_size: 40012277222.86069 - config_name: ga features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 142209710 num_examples: 59156 - name: clean num_bytes: 141702470.68682805 num_examples: 58945 download_size: 121745838 dataset_size: 283912180.686828 - config_name: he features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 4776226234 num_examples: 333874 - name: clean num_bytes: 4760232712.702708 num_examples: 332756 download_size: 3499530576 dataset_size: 9536458946.70271 - config_name: hi features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 1452853579 num_examples: 163093 - name: clean num_bytes: 1443152625.8779714 num_examples: 162004 download_size: 785363639 dataset_size: 2896006204.8779716 - config_name: hr features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 1064630680 num_examples: 202848 - name: clean num_bytes: 1053026432.3195693 num_examples: 200637 download_size: 1028743775 dataset_size: 2117657112.3195693 - config_name: hu features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 3533169653 num_examples: 532427 - name: clean num_bytes: 3510335279.8822336 num_examples: 528986 download_size: 3558613373 dataset_size: 7043504932.882234 - config_name: hy features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 2568868378 num_examples: 303036 - name: clean num_bytes: 2555898405.394963 num_examples: 301506 download_size: 1781142597 dataset_size: 5124766783.394962 - config_name: id features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 2650288629 num_examples: 665622 - name: clean num_bytes: 2630666948.280745 num_examples: 660694 download_size: 2040186206 dataset_size: 5280955577.280745 - config_name: it features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 12188918391 num_examples: 1833639 - name: clean num_bytes: 12163279397.591763 num_examples: 1829782 download_size: 10368836428 dataset_size: 24352197788.591763 - config_name: ja features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 18752888787 num_examples: 1389467 - name: clean num_bytes: 18684866617.717476 num_examples: 1384427 download_size: 15232900753 dataset_size: 37437755404.717476 - config_name: ko features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 3160932689 num_examples: 647897 - name: clean num_bytes: 3151741108.878351 num_examples: 646013 download_size: 3074385022 dataset_size: 6312673797.878351 - config_name: lt features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 781319902 num_examples: 211292 - name: clean num_bytes: 777474168.616436 num_examples: 210252 download_size: 722780874 dataset_size: 1558794070.616436 - config_name: lv features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 531956241 num_examples: 123413 - name: clean num_bytes: 530943303.00615007 num_examples: 123178 download_size: 700342420 dataset_size: 1062899544.00615 - config_name: mr features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 547060763 num_examples: 94133 - name: clean num_bytes: 545450957.3914355 num_examples: 93856 download_size: 278141890 dataset_size: 1092511720.3914356 - config_name: nl features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 6191062892 num_examples: 2135977 - name: clean num_bytes: 6177393712.697661 num_examples: 2131261 download_size: 5179824678 dataset_size: 12368456604.697662 - config_name: 'no' features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 2437191515 num_examples: 617937 - name: clean num_bytes: 2428893175.610127 num_examples: 615833 download_size: 2175299531 dataset_size: 4866084690.6101265 - config_name: pl features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 6864626419 num_examples: 1587721 - name: clean num_bytes: 6861024883.335341 num_examples: 1586888 download_size: 6565864124 dataset_size: 13725651302.335342 - config_name: pt features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 6844185526 num_examples: 1112246 - name: clean num_bytes: 6755821527.2502985 num_examples: 1097886 download_size: 5516209748 dataset_size: 13600007053.250298 - config_name: ro features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 2023493174 num_examples: 442389 - name: clean num_bytes: 2006866635.6197736 num_examples: 438754 download_size: 1652633599 dataset_size: 4030359809.619774 - config_name: ru features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 22550679128 num_examples: 1945063 - name: clean num_bytes: 22439204702.844765 num_examples: 1935448 download_size: 18884603758 dataset_size: 44989883830.844765 - config_name: sa features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 146460109 num_examples: 12156 - name: clean num_bytes: 145435996.68797302 num_examples: 12071 download_size: 95836795 dataset_size: 291896105.687973 - config_name: sk features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 977962245 num_examples: 242235 - name: clean num_bytes: 976048590.4738994 num_examples: 241761 download_size: 1346611201 dataset_size: 1954010835.4738994 - config_name: sl features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 1106532891 num_examples: 183006 - name: clean num_bytes: 1097995332.4385757 num_examples: 181594 download_size: 1006028852 dataset_size: 2204528223.4385757 - config_name: sr features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 3755288114 num_examples: 676605 - name: clean num_bytes: 3735557179.0449376 num_examples: 673050 download_size: 2558022832 dataset_size: 7490845293.044937 - config_name: sv features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 4861956987 num_examples: 2574513 - name: clean num_bytes: 4857071448.365948 num_examples: 2571926 download_size: 3512612936 dataset_size: 9719028435.365948 - config_name: ta features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 1693909025 num_examples: 160651 - name: clean num_bytes: 1682405487.85255 num_examples: 159560 download_size: 985318775 dataset_size: 3376314512.85255 - config_name: te features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 1556095028 num_examples: 87854 - name: clean num_bytes: 1550320823.3066678 num_examples: 87528 download_size: 746686495 dataset_size: 3106415851.306668 - config_name: tr features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 2317236022 num_examples: 534988 - name: clean num_bytes: 2301578085.336879 num_examples: 531373 download_size: 2055444454 dataset_size: 4618814107.336879 - config_name: uk features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 10935662610 num_examples: 1294720 - name: clean num_bytes: 10860532296.947023 num_examples: 1285825 download_size: 8344390939 dataset_size: 21796194906.94702 - config_name: ur features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 918249794 num_examples: 200154 - name: clean num_bytes: 912616078.225986 num_examples: 198926 download_size: 534834968 dataset_size: 1830865872.225986 - config_name: vi features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 3685585608 num_examples: 1288680 - name: clean num_bytes: 3669872935.086358 num_examples: 1283186 download_size: 2646807342 dataset_size: 7355458543.086358 - config_name: zh features: - name: id dtype: string - name: url dtype: string - name: title dtype: string - name: text dtype: string - name: split_text sequence: string splits: - name: train num_bytes: 7820979602 num_examples: 1384748 - name: clean num_bytes: 7781957954.689285 num_examples: 1377839 download_size: 6540517932 dataset_size: 15602937556.689285 configs: - config_name: af data_files: - split: train path: af/train-* - split: clean path: af/clean-* - config_name: ar data_files: - split: train path: ar/train-* - split: clean path: ar/clean-* - config_name: bg data_files: - split: train path: bg/train-* - split: clean path: bg/clean-* - config_name: ca data_files: - split: train path: ca/train-* - split: clean path: ca/clean-* - config_name: cs data_files: - split: train path: cs/train-* - split: clean path: cs/clean-* - config_name: da data_files: - split: train path: da/train-* - split: clean path: da/clean-* - config_name: de data_files: - split: train path: de/train-* - split: clean path: de/clean-* - config_name: el data_files: - split: train path: el/train-* - split: clean path: el/clean-* - config_name: en data_files: - split: train path: en/train-* - split: clean path: en/clean-* - config_name: es data_files: - split: train path: es/train-* - split: clean path: es/clean-* - config_name: et data_files: - split: train path: et/train-* - split: clean path: et/clean-* - config_name: eu data_files: - split: train path: eu/train-* - split: clean path: eu/clean-* - config_name: fa data_files: - split: train path: fa/train-* - split: clean path: fa/clean-* - config_name: fi data_files: - split: train path: fi/train-* - split: clean path: fi/clean-* - config_name: fr data_files: - split: train path: fr/train-* - split: clean path: fr/clean-* - config_name: ga data_files: - split: train path: ga/train-* - split: clean path: ga/clean-* - config_name: he data_files: - split: train path: he/train-* - split: clean path: he/clean-* - config_name: hi data_files: - split: train path: hi/train-* - split: clean path: hi/clean-* - config_name: hr data_files: - split: train path: hr/train-* - split: clean path: hr/clean-* - config_name: hu data_files: - split: train path: hu/train-* - split: clean path: hu/clean-* - config_name: hy data_files: - split: train path: hy/train-* - split: clean path: hy/clean-* - config_name: id data_files: - split: train path: id/train-* - split: clean path: id/clean-* - config_name: it data_files: - split: train path: it/train-* - split: clean path: it/clean-* - config_name: ja data_files: - split: train path: ja/train-* - split: clean path: ja/clean-* - config_name: ko data_files: - split: train path: ko/train-* - split: clean path: ko/clean-* - config_name: lt data_files: - split: train path: lt/train-* - split: clean path: lt/clean-* - config_name: lv data_files: - split: train path: lv/train-* - split: clean path: lv/clean-* - config_name: mr data_files: - split: train path: mr/train-* - split: clean path: mr/clean-* - config_name: nl data_files: - split: train path: nl/train-* - split: clean path: nl/clean-* - config_name: 'no' data_files: - split: train path: no/train-* - split: clean path: no/clean-* - config_name: pl data_files: - split: train path: pl/train-* - split: clean path: pl/clean-* - config_name: pt data_files: - split: train path: pt/train-* - split: clean path: pt/clean-* - config_name: ro data_files: - split: train path: ro/train-* - split: clean path: ro/clean-* - config_name: ru data_files: - split: train path: ru/train-* - split: clean path: ru/clean-* - config_name: sa data_files: - split: train path: sa/train-* - split: clean path: sa/clean-* - config_name: sk data_files: - split: train path: sk/train-* - split: clean path: sk/clean-* - config_name: sl data_files: - split: train path: sl/train-* - split: clean path: sl/clean-* - config_name: sr data_files: - split: train path: sr/train-* - split: clean path: sr/clean-* - config_name: sv data_files: - split: train path: sv/train-* - split: clean path: sv/clean-* - config_name: ta data_files: - split: train path: ta/train-* - split: clean path: ta/clean-* - config_name: te data_files: - split: train path: te/train-* - split: clean path: te/clean-* - config_name: tr data_files: - split: train path: tr/train-* - split: clean path: tr/clean-* - config_name: uk data_files: - split: train path: uk/train-* - split: clean path: uk/clean-* - config_name: ur data_files: - split: train path: ur/train-* - split: clean path: ur/clean-* - config_name: vi data_files: - split: train path: vi/train-* - split: clean path: vi/clean-* - config_name: zh data_files: - split: train path: zh/train-* - split: clean path: zh/clean-* --- # Multilingual Tokenizer Benchmark This dataset includes pre-processed wikipedia data for tokenizer evaluation in [45 languages](https://huggingface.co/datasets/occiglot/tokenizer-wiki-bench/blob/main/README.md#supported-languages). We provide more information on the evaluation task in general [this blogpost](https://occiglot.github.io/occiglot/posts/eu_tokenizer_perfomance/). ## Usage The dataset allows us to easily calculate *tokenizer fertility* and the *proportion of continued words* on any of the supported languages. In the example below we take the Mistral tokenizer and evaluate its performance on Slovak. ```python from transformers import AutoTokenizer from datasets import load_dataset import numpy as np def calculate_metrics(tokens): tmp = np.array([len(y) for y in tokens]) return {'fertility': np.mean(tmp), 'cont_prop': np.count_nonzero(tmp > 1) / tmp.shape[0]} tokenizer_name = 'mistralai/Mistral-7B-v0.1' language = 'sk' #Slovak tokenizer = AutoTokenizer.from_pretrained(tokenizer_name) ds = load_dataset('occiglot/tokenizer-wiki-bench', name=language, split='clean') remove_columns = list(set(ds.column_names) - set(["text"])) ds = ds.map(lambda x: {'tokens': tokenizer(x['split_text'], add_special_tokens=False)['input_ids']} ,num_proc=256, remove_columns=remove_columns, batched=False) remove_columns = None#list(set(ds.column_names)) ds = ds.map(lambda x: calculate_metrics(x['tokens']), num_proc=256, remove_columns=remove_columns, batched=False) df = ds.to_pandas() print('Fertility: ', df.fertility.mean()) print('Prop. continued words:', df.cont_prop.mean()) ``` ## Dataset Creation We loosely follow the approach of [Rust _et al.](https://arxiv.org/abs/2012.15613) using the fast [UDPipe](https://ufal.mff.cuni.cz/udpipe) to pre-split documents into words and subsequently run the tokenizer over isolated words. For all languages we use the respective November 2023 snapshot from [Wikipedia](wikimedia/wikipedia). Since Wikipedia, by nature, contains significantly more numbers and dates than other text and most tokenizers split those into single digits, we filtered all lone-standing numbers from the documents. Additionally, we removed any documents that still contained non-parsed HTML code (less than 1%). ## Licensing We release our curated benchmark and any associated code under [MIT](https://opensource.org/license/mit) license. However, depending on your use case, the licensing conditions of the original [Wikipedia data](https://huggingface.co/datasets/wikimedia/wikipedia#licensing-information) and [UDPipe](https://github.com/ufal/udpipe/tree/udpipe-2?tab=License-1-ov-file) may apply. ## Supported Languages This dataset currently contains pre-processed data for the following languages: | Language | Code | |:-----------|:-------| | Afrikaans | af | | Arabic | ar | | Armenian | hy | | Basque | eu | | Bulgarian | bg | | Catalan | ca | | Croatian | hr | | Czech | cs | | Danish | da | | Dutch | nl | | English | en | | Estonian | et | | Finnish | fi | | French | fr | | German | de | | Greek | el | | Hebrew | he | | Hindi | hi | | Hungarian | hu | | Indonesian | id | | Irish | ga | | Italian | it | | Japanese | ja | | Korean | ko | | Latvian | lv | | Lithuanian | lt | | Marathi | mr | | Norwegian | no | | Persian | fa | | Polish | pl | | Portuguese | pt | | Romanian | ro | | Russian | ru | | Sanskrit | sa | | Serbian | sr | | Slovak | sk | | Slovenian | sl | | Spanish | es | | Swedish | sv | | Tamil | ta | | Telugu | te | | Turkish | tr | | Ukrainian | uk | | Urdu | ur | | Vietnamese | vi |