Datasets:
metadata
task_categories:
- text-generation
language:
- en
- eng
- pt
- por
- es
- spa
- fr
- fra
- zh
- cmn
- ja
- jpn
- de
- deu
- ru
- rus
- ar
- arb
- it
- ita
- pl
- pol
- ko
- kor
- vi
- vie
- code
multilinguality:
- multilingual
pretty_name: Multilingual Tokenizer Benchmark
size_categories:
- 100K<n<1M
configs:
- config_name: default
data_files:
- split: train
path: '**/*.jsonl.gz'
- config_name: en
data_files:
- split: train
path: en/*.jsonl.gz
- config_name: pt
data_files:
- split: train
path: pt/*.jsonl.gz
- config_name: es
data_files:
- split: train
path: es/*.jsonl.gz
- config_name: fr
data_files:
- split: train
path: fr/*.jsonl.gz
- config_name: zh
data_files:
- split: train
path: zh/*.jsonl.gz
- config_name: ja
data_files:
- split: train
path: ja/*.jsonl.gz
- config_name: de
data_files:
- split: train
path: de/*.jsonl.gz
- config_name: ru
data_files:
- split: train
path: ru/*.jsonl.gz
- config_name: ar
data_files:
- split: train
path: ar/*.jsonl.gz
- config_name: it
data_files:
- split: train
path: it/*.jsonl.gz
- config_name: pl
data_files:
- split: train
path: pl/*.jsonl.gz
- config_name: ko
data_files:
- split: train
path: ko/*.jsonl.gz
- config_name: code
data_files:
- split: train
path: code/*.jsonl.gz
- config_name: vi
data_files:
- split: train
path: vi/*.jsonl.gz
source_datasets:
- bigcode/the-stack-smol
- badranx/opus_raw
- HuggingFaceFW/fineweb-2
- sebastiandizon/genius-song-lyrics
- rntc/pubmed_preprocess
- wikimedia/wikipedia
- joelniklaus/Multi_Legal_Pile
- almanach/HALvest
- premio-ai/TheArabicPile_Medical
- ImruQays/Rasaif-Classical-Arabic-English-Parallel-texts
- asas-ai/financial_news
- bigscience-data/roots_ar_ted_talks_iwslt
- premio-ai/TheArabicPile_Lyrics
- AIR-Bench/qa_news_ar
- premio-ai/TheArabicPile_Poetry
- bigscience-data/roots_ar_wikipedia
- bigcode/starcoderdata
- AIR-Bench/qa_news_de
- orionweller/books_mds_incremental
- ElKulako/stocktwits-emoji
- Orion-zhen/dpo-codealpaca-emoji
- Orion-zhen/dpo-mathinstuct-emoji
- Orion-zhen/dpo-physical-reasoning-emoji
- AIR-Bench/qa_finance_en
- bigscience-data/roots_en_ted_talks_iwslt
- orionweller/cc_news_mds_incremental
- DanFosing/public-domain-poetry
- orionweller/arxiv_mds_incremental
- orionweller/pes2o_mds_incremental
- orionweller/reddit_mds_incremental
- HuggingFaceFW/fineweb
- orionweller/wikipedia_mds_incremental
- jorgeortizfuentes/spanish_books
- bigscience-data/roots_es_ted_talks_iwslt
- LeoCordoba/CC-NEWS-ES
- andreamorgar/spanish_poetry
- jorge-henao/disco_poetry_spanish
- eduagarcia/scielo_abstracts
- Arconte/Dominican_reddit_raw_corpus
- bigscience-data/roots_es_wikipedia
- AIR-Bench/qa_finance_fr
- bigscience-data/roots_fr_ted_talks_iwslt
- AIR-Bench/qa_news_fr
- kamarko/ccnews-french-subset
- manu/french_poetry
- bigscience-data/roots_fr_wikipedia
- HiTZ/Multilingual-Medical-Corpus
- globis-university/aozorabunko-clean
- geniacllm/hanrei
- kajuma/CC-news-2024-July-October-cleaned
- p1atdev/modern_haiku
- kunishou/J-ResearchCorpus
- p1atdev/open2ch
- BCCard/BCCard-Finance-Kor-QnA
- jihye-moon/klac_legal_aid_counseling
- joonhok-exo-ai/korean_law_open_data_precedents
- AIR-Bench/qa_news_ko
- daekeun-ml/naver-news-summarization-ko
- werty1248/Korea-Related-Reddit-comments
- werty1248/Korea-Related-Reddit-posts
- WiktorS/polish-news
- winterkitsune/elka-pl-news
- eduagarcia/pd_books_pt
- eduagarcia/LegalPT_dedup
- bigscience-data/roots_pt_ted_talks_iwslt
- eduagarcia/cc_news_pt_v2
- eduagarcia/scielo_papers
- carolina-c4ai/corpus-carolina
- tallesl/quinquilharia
- bigscience-data/roots_pt_wikipedia
- blinoff/medical_qa_ru_data
- sevenreasons/genius-lyrics-russian
- AIR-Bench/qa_news_ru
- IlyaGusev/ru_news
- 0x7o/poemma-10k
- AnyaSchen/russian_poetry_with_keywords
- mlsa-iai-msu-lab/ru_sci_bench
- VietAI/vi_pubmed
- bigscience-data/roots_vi_ted_talks_iwslt
- bigscience-data/roots_vi_binhvq_news_corpus
- bigscience-data/roots_vi_vietnamese_poetry
- DavidLanz/medical_pretrain
- Orion-zhen/dpo-emoji-zh
- Orion-zhen/dpo-ruozhiba-emoji
- shareAI/DPO-zh-en-emoji
- Duxiaoman-DI/FinCorpus
- TigerResearch/tigerbot-law-plugin
- bigscience-data/roots_zh_ted_talks_iwslt
- dirtycomputer/chinese_lyrics
- AIR-Bench/qa_news_zh
- liswei/news-collection-zhtw
- erhwenkuo/poetry-chinese-zhtw
- bigscience-data/roots_zh-cn_wikipedia
Multilingual Tokenizer Benchmark
More details of each subset like word count, character count, original sources, etc, can be found in the dataset_meta.yaml file in the repository root.
Natural language word count functions
Download spacy models
pip install ntlk spacy pygments underthesea camel-tools
python -m spacy download ko_core_news_sm
python -m spacy download ja_core_news_sm
python -m spacy download zh_core_web_sm
import nltk
nltk.download('punkt_tab')
nltk.download('words')
from nltk import word_tokenize as word_tokenize_nltk
from underthesea import word_tokenize as word_tokenize_underthesea
from camel_tools.tokenizers.word import simple_word_tokenize as word_tokenize_camel_tools
import spacy
langs_nltk = {
'en': 'english',
'pt': 'portuguese',
'it': 'italian',
'pl': 'polish',
'de': 'german',
'es': 'spanish',
'fr': 'french',
'ru': 'russian',
}
langs_spacy = {
'ja': spacy.load('ja_core_news_sm', disable=['parser', 'ner', 'lemmatizer', 'tagger', 'attribute_ruler']),
'zh': spacy.load('zh_core_web_sm', disable=['parser', 'ner', 'lemmatizer', 'tagger', 'attribute_ruler']),
'ko': spacy.load('ko_core_news_sm', disable=['parser', 'ner', 'lemmatizer', 'tagger', 'attribute_ruler']),
}
def word_tokenize_spacy(text, model):
doc = model(text)
return [token.text for token in doc if not token.is_space]
def word_tokenize(text, language):
if language in langs_spacy:
return word_tokenize_spacy(text, langs_spacy[language])
elif language in langs_nltk:
return word_tokenize_nltk(text, langs_nltk[language])
elif language == 'vi':
return word_tokenize_underthesea(text)
elif language == 'ar':
return word_tokenize_camel_tools(text)
else:
raise ValueError(f"Language {language} not supported")
#### Usage Example ####
word_count_pt = len(word_tokenize("Olá Mundo.", 'pt'))
word_count_ko = len(word_tokenize("이것은 한국어 문장입니다.", 'ko'))
Code word count function
import nltk
nltk.download('punkt_tab')
from nltk import word_tokenize
from pygments import lex
from pygments.lexers import get_lexer_by_name
from pygments.token import Token
NATURAL_LANGUAGE_PARENT_TOKEN_TYPES = {
Token.Comment,
Token.Literal.String,
Token.Text,
Token.Generic.Heading, # For Markdown,
Token.Generic.Subheading, # For Markdown,
Token.Generic.Emph, # For Markdown *italic*
Token.Generic.Strong, # For Markdown **bold**
Token.Generic.Output, # For shell output,
Token.Generic.Error, # Error messages
Token.Generic.Traceback, # Traceback messages
}
def _is_natural_language_like(tok_type) -> bool:
for nl_parent_type in NATURAL_LANGUAGE_PARENT_TOKEN_TYPES:
if tok_type in nl_parent_type: # This checks if tok_type is nl_parent_type OR a child
return True
return False
def pygments_tokenize(code: str, language: str) -> list[str]:
lexer = get_lexer_by_name(language, stripnl=True, stripall=True)
tokens = lex(code, lexer)
result = []
for tok_type, value in tokens:
if _is_natural_language_like(tok_type):
# Recursively tokenize comment or string content. Use the default english tokenizer.
result.extend(word_tokenize(value))
else:
# Tokenize symbol/keyword/identifier at surface level
result.append(value)
return [t for t in result if t.strip() != '']
#### Usage Example ####
word_count_json = len(pygments_tokenize('{"hello": 1}', 'json'))
word_count_python = len(pygments_tokenize("lambda x: print('hello')", "python"))
Per language stats
| Code | Language | Num. of Domains | Num. of Docs | Num. of Characters | Num. of Words | Num. of Bytes | Data Types |
|---|---|---|---|---|---|---|---|
| code | Programming Languages | 18 | 4260 | 17454895 | 3645230 | 17621590 | C, C++, CSS, Dockerfile, Go, HTML, JSON, Java, JavaScript, Lua, Markdown, PHP, Python, Rust, SQL, Shell, TeX, TypeScript |
| en | English | 14 | 8513 | 14087244 | 2812929 | 14177905 | Biomedical, Books, Emoji Heavy, Finance, Legal, Live Speech, Lyrics, News, Poetry, Scientific Papers, Social Networks, Subtitles, Web, Wikipedia |
| es | Spanish | 12 | 7739 | 12786745 | 2408966 | 13063395 | Biomedical, Books, Legal, Live Speech, Lyrics, News, Poetry, Scientific Papers, Social Networks, Subtitles, Web, Wikipedia |
| fr | French | 11 | 6812 | 11477784 | 2206430 | 11867606 | Biomedical, Finance, Legal, Live Speech, Lyrics, News, Poetry, Scientific Papers, Subtitles, Web, Wikipedia |
| pt | Portuguese | 11 | 7709 | 11546654 | 2207307 | 11887461 | Biomedical, Books, Legal, Live Speech, Lyrics, News, Scientific Papers, Social Networks, Subtitles, Web, Wikipedia |
| zh | Chinese | 11 | 10769 | 3999493 | 2213799 | 10471916 | Biomedical, Emoji Heavy, Finance, Legal, Live Speech, Lyrics, News, Poetry, Subtitles, Web, Wikipedia |
| ar | Arabic | 10 | 12410 | 9909750 | 2005226 | 17685166 | Biomedical, Books, Finance, Live Speech, Lyrics, News, Poetry, Subtitles, Web, Wikipedia |
| ja | Japanese | 10 | 11039 | 3801952 | 2008396 | 9717840 | Books, Legal, Lyrics, News, Poetry, Scientific Papers, Social Networks, Subtitles, Web, Wikipedia |
| de | German | 8 | 5119 | 11016640 | 1841622 | 11207161 | Biomedical, Legal, Lyrics, News, Scientific Papers, Subtitles, Web, Wikipedia |
| ko | Korean | 8 | 9498 | 6060455 | 1602826 | 12532955 | Finance, Legal, Lyrics, News, Social Networks, Subtitles, Web, Wikipedia |
| ru | Russian | 8 | 8729 | 9105314 | 1612537 | 16277665 | Biomedical, Lyrics, News, Poetry, Scientific Papers, Subtitles, Web, Wikipedia |
| vi | Vietnamese | 8 | 8973 | 8533015 | 1602500 | 10911187 | Biomedical, Live Speech, Lyrics, News, Poetry, Subtitles, Web, Wikipedia |
| it | Italian | 7 | 6976 | 7706475 | 1408363 | 7792425 | Biomedical, Legal, Lyrics, Scientific Papers, Subtitles, Web, Wikipedia |
| pl | Polish | 6 | 3990 | 7278038 | 1204690 | 7694868 | Legal, Lyrics, News, Subtitles, Web, Wikipedia |