Hugging Face's logo

Datasets:
vpermilp
/
spellman

Tasks:
Text Classification
Modalities:
Text
Formats:
parquet
Languages:
Bashkir
Belarusian
Bulgarian
Size:
1M - 10M
Tags:
language-identification
cyrillic
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
spellman / README.md
vpermilp's picture
vpermilp
v16 mix: 4,208,639 rows — twin gates, short Russian re-sourced, referees held out
1b9641e verified
preview code
|
Raw History Blame Contribute Delete
21.2 kB
metadata 
language:
  - bak
  - bel
  - bul
  - che
  - chv
  - deu
  - eng
  - fra
  - kaz
  - kir
  - kpv
  - mhr
  - mkd
  - mon
  - oss
  - por
  - rus
  - sah
  - spa
  - srp
  - tat
  - tgk
  - tyv
  - udm
  - ukr
  - uzn
license: other
license_name: mixed-upstream-per-source-in-manifest
task_categories:
  - text-classification
tags:
  - language-identification
  - cyrillic
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet
      - split: val
        path: data/val-*.parquet
      - split: test
        path: data/test-*.parquet
dataset_info:
  features:
    - name: lang
      dtype: string
    - name: text
      dtype: string
  splits:
    - name: train
      num_examples: 2515126
      num_bytes: 185809059
    - name: val
      num_examples: 974866
      num_bytes: 86374024
    - name: test
      num_examples: 718647
      num_bytes: 68622397

spellman — Cyrillic language-ID training mix

Training dataset for the spellman detector: 26 languages (21 Cyrillic-column + the 5 Latin big-5 — the model's other 4 classes are script-routed and never appear as rows), one row per sample (lang, text), content-addressed train/val/test split.

License: mixed — each row inherits its upstream corpus's license; the per-source recipe (and therefore attribution) is recorded in manifest.json. No upstream is non-commercial (NC): the 2026-08 license audit removed/replaced every NC-licensed or NC-derived source. Upstreams whose cards state no license are listed as unstated below and are tracked by a follow-up audit wave.

Upstream licenses per source (HF card metadata + the 2026-08 audit):

source license
fineweb2: ODC-By-1.0
tatoeba: CC-BY-2.0-FR (attribution embedded per sentence)
hf:cis-lmu/Glot500 aggregate, per-source — NC/restricted slices row-filtered (Leipzig*/nllb_other_til/mtdata/Bible*); uzb_Cyrl restricted to the Earthlings slice (GPL-3)
hf:cis-lmu/Glot500 aggregate, per-source — NC/restricted slices row-filtered (Leipzig*/nllb_other_til/mtdata/Bible*); uzb_Cyrl restricted to the Earthlings slice (GPL-3)
hf:cis-lmu/Glot500 aggregate, per-source — NC/restricted slices row-filtered (Leipzig*/nllb_other_til/mtdata/Bible*); uzb_Cyrl restricted to the Earthlings slice (GPL-3)
hf:cis-lmu/Glot500 aggregate, per-source — NC/restricted slices row-filtered (Leipzig*/nllb_other_til/mtdata/Bible*); uzb_Cyrl restricted to the Earthlings slice (GPL-3)
hf:cis-lmu/Glot500 aggregate, per-source — NC/restricted slices row-filtered (Leipzig*/nllb_other_til/mtdata/Bible*); uzb_Cyrl restricted to the Earthlings slice (GPL-3)
hf:ailabykt/sakha-corpus-mono CC-BY-4.0
hf:wikimedia/wikipedia CC-BY-SA-3.0 + GFDL (dual, text)
hf:AigizK/tatar-russian-parallel-corpora unstated on card
hf:HuggingFaceFW/fineweb-2 ODC-By-1.0
hf:AigizK/bashkir-russian-parallel-corpora CC-BY-4.0
hf:Agisight/tyv-rus-200k CC-BY-4.0
hf:ai-forever/udmurt-corpora MIT
hf:udmurtNLP/zerpal unstated on card
hf:d0rj/ru-mhr-parallel CC-BY-4.0
opus:translatewiki CC-BY-SA-3.0+ (translatewiki.net content)
hf:NM-development/nmd-ce-ru-171k-v0 MIT
csv:rusentitweet_train.csv unstated (research corpus; audit wave 2)
jsonl:seeds/hard_negatives.jsonl inherits FineWeb-2 (model-labeled _removed windows)
jsonl:cache/hard_negatives_scaled.jsonl inherits FineWeb-2 (model-labeled _removed windows)
ukr_tweets: unstated (github.com/saganoren/ukr-twi-corpus; audit wave 2)
mn_social: unstated (github.com/ganaxy/diploma; audit wave 2)
kazsandra: unstated (github.com/IS2AI/KazSAnDRA; audit wave 2)
hf:alexantonov/chuvash_mono CC0-1.0
hf:alexantonov/chuvash_russian_parallel CC0-1.0
hf:tahrirchi/uz-crawl Apache-2.0
hf:averoo/sakha-oscar unstated on card
hf:allenai/MADLAD-400 ODC-By-1.0
hf:allenai/MADLAD-400 ODC-By-1.0
hf:YShynkarov/COSMUS MIT
hf:maaxap/BelarusianGLUE Apache-2.0
hf:DGurgurov/macedonian_sa MIT
hf:DGurgurov/bulgarian_sa MIT
hf:cardiffnlp/tweet_sentiment_multilingual unstated on card
hf:cardiffnlp/tweet_sentiment_multilingual unstated on card
hf:cardiffnlp/tweet_sentiment_multilingual unstated on card
hf:cardiffnlp/tweet_sentiment_multilingual unstated on card
hf:cardiffnlp/tweet_sentiment_multilingual unstated on card
hf:contemmcm/sentiment140 unstated on card
hf:FrancophonIA/french_tweets unstated on card
hf:fpaulino/portuguese-tweets unstated on card
hf:mteb/told-br CC-BY-SA-4.0
hf:NLP-UniBW/tweets_about_german_politicians_jan_feb_2025 unstated on card
hf:pysentimiento/spanish-tweets unstated on card
ukr_tweets: unstated (github.com/saganoren/ukr-twi-corpus; audit wave 2)
kazsandra: unstated (github.com/IS2AI/KazSAnDRA; audit wave 2)
mn_social: unstated (github.com/ganaxy/diploma; audit wave 2)
hf:cardiffnlp/tweet_sentiment_multilingual unstated on card
hf:cardiffnlp/tweet_sentiment_multilingual unstated on card
hf:cardiffnlp/tweet_sentiment_multilingual unstated on card
hf:cardiffnlp/tweet_sentiment_multilingual unstated on card
hf:cardiffnlp/tweet_sentiment_multilingual unstated on card
hf:mteb/told-br CC-BY-SA-4.0
hf:FrancophonIA/french_tweets unstated on card
hf:contemmcm/sentiment140 unstated on card
hf:DGurgurov/macedonian_sa MIT
hf:DGurgurov/bulgarian_sa MIT
diverse:HuggingFaceFW/fineweb-2 ODC-By-1.0
diverse:HuggingFaceFW/fineweb-2 ODC-By-1.0
csv:rusentitweet_train.csv unstated (research corpus; audit wave 2)
jsonl:seeds/relabel_rus.jsonl unstated (derived seed set; audit wave 2)
diverse:cache/wikisource-5da9d07b30.jsonl CC-BY-SA-4.0 (Wikisource text)
diverse:cache/kazsandra-eb6571e000.jsonl unstated (github.com/IS2AI/KazSAnDRA; audit wave 2)
diverse:HPLT/HPLT2.0_cleaned CC0-1.0
diverse:cache/hf-b407fce95d.jsonl MIT
diverse:cache/hf-1f351fe715.jsonl Apache-2.0
diverse:cache/hf-c2a1d6fbcb.jsonl unstated on card
diverse:cache/hf-b242f3ef58.jsonl CC-BY-4.0
diverse:cache/hf-7e962c6753.jsonl CC0-1.0
diverse:cache/hf-0bea6ddfdd.jsonl CC-BY-4.0
diverse:cache/hf-cd00eaba11.jsonl aggregate, per-source — NC/restricted slices row-filtered (Leipzig*/nllb_other_til/mtdata/Bible*); uzb_Cyrl restricted to the Earthlings slice (GPL-3)
hf:allenai/MADLAD-400 ODC-By-1.0
hf:allenai/MADLAD-400 ODC-By-1.0
hf:wikimedia/wikipedia CC-BY-SA-3.0 + GFDL (dual, text)
hf:HPLT/HPLT2.0_cleaned CC0-1.0
hf:allenai/MADLAD-400 ODC-By-1.0
hf:wikimedia/wikipedia CC-BY-SA-3.0 + GFDL (dual, text)
opus:GoURMET CC0 (OPUS release)
hf:HPLT/HPLT2.0_cleaned CC0-1.0
hf:alifbank/Tajik MIT
hf:allenai/MADLAD-400 ODC-By-1.0
hf:wikimedia/wikipedia CC-BY-SA-3.0 + GFDL (dual, text)
fineweb2: ODC-By-1.0
hf:tahrirchi/uz-books-v2 MIT
hf:cis-lmu/Glot500 aggregate, per-source — NC/restricted slices row-filtered (Leipzig*/nllb_other_til/mtdata/Bible*); uzb_Cyrl restricted to the Earthlings slice (GPL-3)
hf:vpermilp/lid-sah MIT (compilation; row texts from cited public sources)
hf:vpermilp/lid-sah MIT (compilation; row texts from cited public sources)
hf:vpermilp/lid-tyv MIT (compilation; row texts from cited public sources)
hf:vpermilp/lid-tyv MIT (compilation; row texts from cited public sources)
hf:vpermilp/lid-kpv MIT (compilation; row texts from cited public sources)
hf:vpermilp/lid-kpv MIT (compilation; row texts from cited public sources)
hf:vpermilp/lid-mhr MIT (compilation; row texts from cited public sources)
hf:vpermilp/lid-mhr MIT (compilation; row texts from cited public sources)
hf:vpermilp/lid-oss MIT (compilation; row texts from cited public sources)
hf:vpermilp/lid-oss MIT (compilation; row texts from cited public sources)
hf:vpermilp/lid-udm MIT (compilation; row texts from cited public sources)
ukr_tweets: unstated (github.com/saganoren/ukr-twi-corpus; audit wave 2)
hf:Den4ikAI/russian_dialogues MIT
hf:Den4ikAI/russian_dialogues MIT
hf:hausmer/dvach_chat other (uploader's label) — a member's export of the Dvach Telegram group chat; message text only, author names not included
hf:nyuuzyou/ruforum CC0-1.0

Recipe (also recorded, byte-exact, in manifest.json):

  • seed 42, cap_per_lang=120000
  • augmentation: wild 0.3, short 0.2 (train/val only; test pristine)
  • short floor 0.5
  • orthographic twin gate 0.5: rows whose spelling is mostly a twin language of their label are dropped (e.g. Ukrainian tweets tagged ru)
  • lexical twin gate 4.0 nats: rows whose words are a twin language's are dropped (short Ukrainian with no Ukrainian-only letter)
  • referee holdout: rows whose text appears in an evaluation referee (tatoeba_eval.tsv, rusentitweet_eval_v2.tsv, cosmus_rus_eval.tsv, short_eval.tsv, lit_rus_eval.tsv) are in no split

Sources in mix order (dedup is first-source-wins):

source options
fineweb2 docs_per_lang=3600, per_doc=4, langs_exclude='tgk'
tatoeba train_per_lang=8000
hf repo='cis-lmu/Glot500', config='tat_Cyrl', exclude='dataset=Leipzig*
hf repo='cis-lmu/Glot500', config='tgk_Cyrl', lang='tgk', exclude='dataset=Leipzig*
hf repo='cis-lmu/Glot500', config='sah_Cyrl', exclude='dataset=Leipzig*
hf repo='cis-lmu/Glot500', config='udm_Cyrl', exclude='dataset=Leipzig*
hf repo='cis-lmu/Glot500', config='tyv_Cyrl', exclude='dataset=Leipzig*
hf repo='ailabykt/sakha-corpus-mono', lang='sah', docs=8000, per_doc=3
hf repo='wikimedia/wikipedia', config='20231101.tt', lang='tat', docs=2000, per_doc=4
hf repo='AigizK/tatar-russian-parallel-corpora', column='tat', lang='tat', docs=999999, per_doc=2, streaming=False
hf repo='HuggingFaceFW/fineweb-2', config='tat_Latn', lang='tat', docs=10000, per_doc=3
hf repo='AigizK/bashkir-russian-parallel-corpora', column='ba', lang='bak', docs=30000, per_doc=2
hf repo='Agisight/tyv-rus-200k', column='tyv', lang='tyv', docs=30000, per_doc=2
hf repo='ai-forever/udmurt-corpora', lang='udm', docs=30000, per_doc=2
hf repo='udmurtNLP/zerpal', column='string', lang='udm', docs=20000, per_doc=3
hf repo='d0rj/ru-mhr-parallel', column='mhr', lang='mhr', docs=30000, per_doc=2
opus corpus='translatewiki', src='ce', tgt='en', lang='che'
hf repo='NM-development/nmd-ce-ru-171k-v0', column='ce', lang='che', docs=999999, per_doc=1, streaming=False
csv path='rusentitweet_train.csv', column='text', lang='rus'
jsonl path='seeds/hard_negatives.jsonl'
jsonl path='cache/hard_negatives_scaled.jsonl'
ukr_tweets limit=50000
mn_social defaults
kazsandra defaults
hf repo='alexantonov/chuvash_mono', column='chv', lang='chv', raw=True, docs=500000, max_chars=512
hf repo='alexantonov/chuvash_russian_parallel', column='chv', lang='chv', raw=True, docs=300000, max_chars=512
hf repo='tahrirchi/uz-crawl', files='data/telegram_blogs*', lang='uzn', raw=True, cyr=0.6, docs=400000, max_chars=512
hf repo='averoo/sakha-oscar', lang='sah', raw=True, docs=0, streaming=False, max_chars=512
hf repo='allenai/MADLAD-400', files='data/tyv/tyv_clean_0000.jsonl.gz', lang='tyv', raw=True, docs=0, streaming=False, max_chars=512
hf repo='allenai/MADLAD-400', files='data/sah/sah_clean_0000.jsonl.gz', lang='sah', raw=True, docs=0, streaming=False, max_chars=512
hf repo='YShynkarov/COSMUS', column='document_content', where='language_manual=ukrainian', lang='ukr', raw=True, docs=0, streaming=False, max_chars=512
hf repo='maaxap/BelarusianGLUE', config='besls', column='sentence', lang='bel', raw=True, docs=0, streaming=False
hf repo='DGurgurov/macedonian_sa', column='text', lang='mkd', raw=True, docs=0, streaming=False, max_chars=512
hf repo='DGurgurov/bulgarian_sa', column='text', lang='bul', raw=True, docs=0, streaming=False, max_chars=512
hf repo='cardiffnlp/tweet_sentiment_multilingual', files='data/english/train.jsonl', lang='eng', raw=True, docs=0, streaming=False
hf repo='cardiffnlp/tweet_sentiment_multilingual', files='data/french/train.jsonl', lang='fra', raw=True, docs=0, streaming=False
hf repo='cardiffnlp/tweet_sentiment_multilingual', files='data/german/train.jsonl', lang='deu', raw=True, docs=0, streaming=False
hf repo='cardiffnlp/tweet_sentiment_multilingual', files='data/portuguese/train.jsonl', lang='por', raw=True, docs=0, streaming=False
hf repo='cardiffnlp/tweet_sentiment_multilingual', files='data/spanish/train.jsonl', lang='spa', raw=True, docs=0, streaming=False
hf repo='contemmcm/sentiment140', column='text', lang='eng', raw=True, split='complete', docs=400000, max_chars=512
hf repo='FrancophonIA/french_tweets', lang='fra', raw=True, docs=0, streaming=False, max_chars=512
hf repo='fpaulino/portuguese-tweets', column='tweet_text', lang='por', raw=True, docs=120000, max_chars=512
hf repo='mteb/told-br', lang='por', raw=True, drop_cjk=True, docs=0, streaming=False, max_chars=512
hf repo='NLP-UniBW/tweets_about_german_politicians_jan_feb_2025', where='language=de', lang='deu', raw=True, docs=0, max_chars=512
hf repo='pysentimiento/spanish-tweets', files='data/train-00000-of-00166-*.parquet', lang='spa', raw=True, docs=400000, max_chars=512
ukr_tweets lang='ukr', twitter_lang='uk', min_chars=3, max_chars=19, cyr=0.2, limit=200000
kazsandra min_chars=3, max_chars=19, cyr=0.2
mn_social min_chars=3, max_chars=19
hf repo='cardiffnlp/tweet_sentiment_multilingual', files='data/english/train.jsonl', lang='eng', raw=True, docs=0, streaming=False, min_chars=3, max_chars=19
hf repo='cardiffnlp/tweet_sentiment_multilingual', files='data/french/train.jsonl', lang='fra', raw=True, docs=0, streaming=False, min_chars=3, max_chars=19
hf repo='cardiffnlp/tweet_sentiment_multilingual', files='data/portuguese/train.jsonl', lang='por', raw=True, docs=0, streaming=False, min_chars=3, max_chars=19
hf repo='cardiffnlp/tweet_sentiment_multilingual', files='data/german/train.jsonl', lang='deu', raw=True, docs=0, streaming=False, min_chars=3, max_chars=19
hf repo='cardiffnlp/tweet_sentiment_multilingual', files='data/spanish/train.jsonl', lang='spa', raw=True, docs=0, streaming=False, min_chars=3, max_chars=19
hf repo='mteb/told-br', lang='por', raw=True, drop_cjk=True, docs=0, streaming=False, min_chars=3, max_chars=19
hf repo='FrancophonIA/french_tweets', lang='fra', raw=True, docs=0, streaming=False, min_chars=3, max_chars=19
hf repo='contemmcm/sentiment140', column='text', lang='eng', raw=True, split='complete', docs=400000, min_chars=3, max_chars=19
hf repo='DGurgurov/macedonian_sa', column='text', lang='mkd', raw=True, docs=0, streaming=False, min_chars=3, max_chars=19
hf repo='DGurgurov/bulgarian_sa', column='text', lang='bul', raw=True, docs=0, streaming=False, min_chars=3, max_chars=19
diverse lang='rus', pool_repo='HuggingFaceFW/fineweb-2', pool_config='rus_Cyrl', pool_docs=300000, budget=20000
diverse lang='ukr', pool_repo='HuggingFaceFW/fineweb-2', pool_config='ukr_Cyrl', pool_docs=300000, budget=20000
csv path='rusentitweet_train.csv', column='text', lang='rus', min_chars=3, max_chars=19
jsonl path='seeds/relabel_rus.jsonl'
diverse lang='rus', pool_file='cache/wikisource-5da9d07b30.jsonl', budget=12000
diverse lang='kaz', pool_file='cache/kazsandra-eb6571e000.jsonl', budget=16000
diverse lang='kir', pool_repo='HPLT/HPLT2.0_cleaned', pool_config='kir_Cyrl', pool_docs=200000, budget=16000, pool_no_chars='ҕһ'
diverse lang='tgk', pool_file='cache/hf-b407fce95d.jsonl', budget=16000
diverse lang='uzn', pool_file='cache/hf-1f351fe715.jsonl', budget=16000
diverse lang='tat', pool_file='cache/hf-c2a1d6fbcb.jsonl', budget=16000
diverse lang='bak', pool_file='cache/hf-b242f3ef58.jsonl', budget=16000
diverse lang='chv', pool_file='cache/hf-7e962c6753.jsonl', budget=16000
diverse lang='sah', pool_file='cache/hf-0bea6ddfdd.jsonl', budget=16000
diverse lang='tyv', pool_file='cache/hf-cd00eaba11.jsonl', budget=16000
hf repo='allenai/MADLAD-400', files='data/ce/ce_clean_0000.jsonl.gz', lang='che', raw=True, docs=0, streaming=False, max_chars=512, cyr=0.6
hf repo='allenai/MADLAD-400', files='data/ce/ce_noisy_0000.jsonl.gz', lang='che', raw=True, docs=0, streaming=False, max_chars=512, cyr=0.6
hf repo='wikimedia/wikipedia', config='20231101.ce', lang='che', docs=30000, per_doc=3
hf repo='HPLT/HPLT2.0_cleaned', config='kir_Cyrl', lang='kir', docs=30000, per_doc=4, no_chars='ҕһ'
hf repo='allenai/MADLAD-400', files='data/ky/ky_clean_0000.jsonl.gz', lang='kir', raw=True, docs=0, streaming=False, max_chars=512, cyr=0.6
hf repo='wikimedia/wikipedia', config='20231101.ky', lang='kir', docs=10000, per_doc=4
opus corpus='GoURMET', src='ky', tgt='ru', lang='kir'
hf repo='HPLT/HPLT2.0_cleaned', config='tgk_Cyrl', lang='tgk', docs=30000, per_doc=4, no_chars='ўы'
hf repo='alifbank/Tajik', files='tg_sentences.txt', lang='tgk', raw=True, docs=0, streaming=False, min_chars=20, max_chars=512
hf repo='allenai/MADLAD-400', files='data/tg/tg_clean_0000.jsonl.gz', lang='tgk', raw=True, docs=0, streaming=False, max_chars=512, cyr=0.6, no_chars='ўы'
hf repo='wikimedia/wikipedia', config='20231101.tg', lang='tgk', docs=10000, per_doc=4
fineweb2 docs_per_lang=3600, per_doc=4, langs='tgk', no_chars='ўы'
hf repo='tahrirchi/uz-books-v2', files='data/cyr-*.parquet', lang='uzn', docs=30000, per_doc=4
hf repo='cis-lmu/Glot500', config='uzb_Cyrl', lang='uzn', raw=True, docs=200000, max_chars=512, cyr=0.6, where='dataset=Earthlings'
hf repo='vpermilp/lid-sah', config='moderated', lang='sah', raw=True, docs=0, streaming=False, max_chars=512
hf repo='vpermilp/lid-sah', config='wild', lang='sah', raw=True, docs=0, streaming=False, max_chars=512
hf repo='vpermilp/lid-tyv', config='moderated', lang='tyv', raw=True, docs=0, streaming=False, max_chars=512
hf repo='vpermilp/lid-tyv', config='wild', lang='tyv', raw=True, docs=0, streaming=False, max_chars=512
hf repo='vpermilp/lid-kpv', config='moderated', lang='kpv', raw=True, docs=0, streaming=False, max_chars=512
hf repo='vpermilp/lid-kpv', config='wild', lang='kpv', raw=True, docs=0, streaming=False, max_chars=512
hf repo='vpermilp/lid-mhr', config='moderated', lang='mhr', raw=True, docs=0, streaming=False, max_chars=512
hf repo='vpermilp/lid-mhr', config='wild', lang='mhr', raw=True, docs=0, streaming=False, max_chars=512
hf repo='vpermilp/lid-oss', config='moderated', lang='oss', raw=True, docs=0, streaming=False, max_chars=512
hf repo='vpermilp/lid-oss', config='wild', lang='oss', raw=True, docs=0, streaming=False, max_chars=512
hf repo='vpermilp/lid-udm', config='wild', lang='udm', raw=True, docs=0, streaming=False, max_chars=512
ukr_tweets lang='rus', twitter_lang='ru', min_chars=3, max_chars=19, cyr=0.2, limit=100000, lex_own=3
hf repo='Den4ikAI/russian_dialogues', column='question', docs=0, streaming=False, sample=6000, lang='rus', raw=True, min_chars=3, max_chars=19, drop_long=True, cyr=0.6, no_chars='іїєґўәғқңөұүһІЇЄҐЎӘҒҚҢӨҰҮҺ'
hf repo='Den4ikAI/russian_dialogues', column='answer', docs=0, streaming=False, sample=6000, lang='rus', raw=True, min_chars=3, max_chars=19, drop_long=True, cyr=0.6, no_chars='іїєґўәғқңөұүһІЇЄҐЎӘҒҚҢӨҰҮҺ'
hf repo='hausmer/dvach_chat', docs=0, streaming=False, sample=12000, lang='rus', raw=True, min_chars=3, max_chars=19, drop_long=True, cyr=0.6, no_chars='іїєґўәғқңөұүһІЇЄҐЎӘҒҚҢӨҰҮҺ'
hf repo='nyuuzyou/ruforum', docs=1500000, sample=12000, lang='rus', raw=True, min_chars=3, max_chars=19, drop_long=True, cyr=0.6, no_chars='іїєґўәғқңөұүһІЇЄҐЎӘҒҚҢӨҰҮҺ'

The exact spec strings, including option quoting, are in manifest.json (argv).