Skip to content

Latest commit

 

History

History
68 lines (67 loc) · 76.5 KB

File metadata and controls

68 lines (67 loc) · 76.5 KB

Sources

USED Dataset Name keys Reliable Language Tags Domain Related Links License
Yes, (V1, V2, V3) Wikipedia wikipedia, wili2018, leipzigwiki Yes, however! Articles HF_Wikipedia dumps.wikipedia, wili_2018, Leipzig CC BY-SA 4.0
Yes, (V1, V2, V3) SETIMES setimes Yes News Opus_SETIMES CC-BY-SA 3.0
Yes, (V1, V2, V3) Tatoeba tatoeba Yes, mostly Crowdsourcing Tatoeba CC-BY
Yes, (V1, V2, V3) Global Voices globalvoices Yes News stories Opus_GlobalVoices CC BY 3.0
Yes, (V1, V2, V3) XL-Sum xlsum Yes BBC News HF_xlsum, Github_xlsum CC BY-NC-SA 4.0
Yes, (V1, V2, V3) Leipzig - News and Newscrawl leipzignews Yes News Leipzig CC BY-NC-SA 3.0
Yes, (V3) FLORES dev floresdev Yes, mostly Human Translation FLORES CC-BY-SA 4.0
Yes, (V1, V2, metadata check for V3) NLLB_seed nllbseed Yes Professionally-translated sentences (Wikipedia domain) NLLB_Seed CC-BY-SA 4.0
Yes, (V1, V2, clean for V3) MT-560v1 mt560 ?, openlid version is much cleaner. Multiple domain OpenLID compilation Apache License 2.0
Yes, (V1, V2, V3) Autshumato autshumato Yes Government domain HF_autshumato CC Attribution 2.5 South Africa License
Yes, (V1, V2, metadata check for V3) Open Bibles pbc, cbc, bibleslti Yes, but some closely related languages (or similar name) might be confused. Bible versions 1000Langs, PBC, CorpusCrawler, PNG, Open-Bibles, Bible.is, ebible, biblenlp-corpus, JHUBC, bible.com Mixed, CC family mostly CC BY-NC-ND
Yes (partly), (V1, V2, more clean for V3) JW jw, jw300, jw300data Yes, except sign language codes New World Bible Masakhane-Mt, JW, JW300 Usage for your own personal and non-commercial purposes is permitted. Copyrighted and distribution is not allowed.
Yes (partly), (V1, V2, V3) LTI lti ? Multiple domain whatlang Custom License/ Open partly
Yes (partly), (V1, V2, delete some languages for V3) Arabic (DART, SHAMI, TSAC, PADIC, AOC, Arabic Dialects Dataset, MADAR) iadd, madar, arabdialects ? Multiple domain IADD, MADAR, Arabic Dialects Multiple open licenses
Yes, (V1, V2, V3) Persian (TEP, MIZAN) tep, mizan Yes literature, subtitle TEP, MIZAN TEP: GNU General Public License, MIZAN: CC BY 4.0
Yes, (V1, V2, V3) TIL Corpus til ? Multiple domain TIL-MT CC BY-NC-SA 4.0
Yes, (V1, V2, V3) bho-resources bhojpuri Yes News bho-resources CC BY-NC-SA 4.0
Yes, (V1, V2, V3) Guaraní Parallel Set guarani Yes News Guaraní Parallel Set No explicit license
Yes, (V1, V2, V3) HKCanCor hkcancor Yes Transcribed conversations hkcancor CC BY 4.0
Yes, (V2, V3) ai4d challenge (Nyanja) ai4d Yes News ai4d-malawi-news No explicit license
Yes, (V2, V3) Wanca 2016 wanca2016 ? Web wanca2016 CC - BY
No, (V2, delete for V3) smugri No, after training V2 model we find this data is not clean News smugri-data CC BY 4.0
No, (V2, delete for V3) finno-ugric ? ? finno-ugric-train CC BY 4.0
Yes, (V2, V3) smugri-flores smugriflores Yes Human Translation smugri-flores-testset CC BY 4.0
Yes, (V2, V3) Abkhaz National Corpus abnc Yes grammatically annotated text (linguistics, literary studies, history, political and social sciences) Abkhaz National Corpus, abkhaz_text Public domain (cc0-1.0)
Yes, (V2, V3) Luo News (Radio Ramogi) radioramogi Yes News Luo CC BY 4.0
Yes, (V2, more metadata check for V3) Lyrics lyrics Yes, but there might be some issues with the translation part. Song lyrics lyricstranslate Copyright issues prevent distributing the original lyrics. Lyrics on Lyricstranslate.com are licensed through Musixmatch. Translations on Lyricstranslate.com belong to their authors.
Yes, (V2, V3) GlotSparse glotsparse Yes News and Articles HF_GlotSparse, Github_GlotSparse Web crawl data; copyright retained by original content owners.
Yes, (V2, more metadata check for V3) GlotStoryBook glotstorybook Yes StoryBooks HF_GlotStoryBook, Github_GlotStoryBook mostly CC family
Yes (partly), (V2, more clean for V3) Universal Dependencies v2.12 ud Yes, mostly Multiple domain UD CC family
Yes, (V2, V3) CommonVoice v11 commonvoice Yes, mostly crowdsourcing CommonVoice v11 Public domain (cc0-1.0)
Yes, (V2, V3) GOUD.MA goud ?, only the headlines are definitely written in Moroccan Darija, but some noises exist. News Goud-sum No explicit license
Yes, (V2, V3) Vuk'uzenzele vukuzenzele Yes government domain vukuzenzele-monolingual License for Data - CC BY 4.0
Yes, (V2, V3) Masakhanews masakhanews Yes News masakhane/masakhanews CC BY-NC 4.0
Yes, (V2, V3) AfriQA afriqa Yes Human Translation masakhane/afriqa CC BY-NC 4.0
Todo African News Corpus Yes News African News Corpus Non-Commercial Government Licence
Todo AfriSenti ?, location tags converted to language tags followed by annotation Tweets AfriSenti CC BY 4.0
Todo Bambara Dataset - Sentiment ?, followed by annotation CommonCrawl Bambara Dataset No explicit license
Todo TUNIZI Dataset - Sentiment ?, followed by annotation YouTube videos comments TUNIZI Dataset, HF_TUNIZI No explicit license
Todo CTAB ?, followed by manual verification Facebook Public Pages Zenodo_CTAB CC BY 4.0
Todo Open Subtitles ? Movie subtitles OpenSubtitles
Todo GNOME Yes Human Translation opus_GNOME, HF_opus_GNOME, gnome/releases
Todo KDE4 Yes Human Translation opus_kde4 HF_kde4 kde4
Todo Ubuntu Yes Human Translation HF_opus_ubuntu
Todo Web Inventory of Transcribed & Translated (WIT) Ted Talks ? Ted talks ted_talks_iwslt
Todo igbo_monolingual Yes News, Radio, Books igbo_monolingual
Todo QADI ?, location tags can be converted to language tags, but not very reliable for LID. Tweets QADI Tweet IDs Apache License 2.0
Todo Mot Yes VOA News mot MIT License
Todo gov-za Yes government domain gov-za-monolingual License for Data - CC BY 4.0
Yes, (V3) NusaT, NusaP, NusaX nusa Yes Human Translation indonlp, nusa-writes, NusaX CC-BY-SA 4.0 or Apache License 2.0
Yes, (V3) MT Shared Task American NLP americasnlp2024 americasnlp2024
Yes, (V3) ShaShiYaYi shashiyayi multilingual-data-peru
No, (delete for v3, it's code-switch) Hinglish collection english-to-hinglish
Yes, (V3) Indic corpora in22conv, in22gen In22Conv, In22Gen CC-BY-SA 4.0
Yes, (V3) Yue parallel text yuecmnengparallel yue-cmn-eng
Todo Dakshina dakshina
Yes, (V3) Bhasha-Abhijnaanam bhashaabhijnaanam Bhasha-Abhijnaanam
Yes, (V3) Bloom Library bloombooks bloom-lm CC family 4.0
Yes, (V4) FineWeb2 fineweb2 only und labels after recycling with GlotScript Web crawl / CommonCrawl derived web text FineWeb2 Web crawl data; copyright retained by original content owners.
Yes, (V4) Unicode/Google Fonts language samples googlefonts, unicodefonts only very low resource scripts Font samples / multilingual text examples rolandwalker/unicode-fonts , languagegeek.com
Yes, (V4) MINEDU Repository (Ministerio de Educación del Perú) minedu Yes (government datasets) Government / Education / Indigenous language resources repositorio.minedu , https://datos.minedu.gob.pe/ Peruvian Government Open Data License
Yes, (V4) GlotWeb / GlotCC corpus glotweb Yes, only confimred one Web pages for low-resource scripts and languages GlotWeb , GlotCC-V1 Web crawl data; copyright retained by original content owners.
Yes, (V4) Wiktionary wordlists wiktionary Yes Dictionary / lexical entries wiktionary_extract CC BY-SA 3.0 (Wiktionary content)
Yes, (V4) UDHR (only for rare scripts), the rest is for LID benchmarking udhr Yes Universal Declaration of Human Rights Omniglot, UDHR-LID Public Domain (United Nations)