Datasets:
African Languages Lab Multi-Open
This dataset is an open subset of African-Languages-Lab/all-lab-text-multi.
It contains only multilingual examples whose source/provenance (dataset_name) matches an
approved OPUS / open corpus. It is intended for multilingual African-language NLP research,
especially machine translation and multilingual language modeling.
This is NOT the full All Lab multilingual dataset. The full All Lab data combines open-source corpora, All Voices contributions, and community-partner data. This
multi-openrelease intentionally keeps only the OPUS / open-source-derived rows — All Voices, community, partner, internal, and unknown-provenance rows are excluded.
- Source repo:
African-Languages-Lab/all-lab-text-multi - This (open) repo:
African-Languages-Lab/multi-open - Layout: one config per language pair (e.g.
english-swahili), mirroring the source — each config keeps the exact same columns as the source (the two language text columns, their token counts,dataset_name, andsource_data_url). Nothing is renamed or dropped. - Filtering rule: a row is kept only if its
dataset_namematches an approved OPUS/open corpus (filename/<corpus> <lang-pair>style labels handled; generic look-alikes excluded).
Dataset statistics
- Total rows: 323,148,690
- Language pairs (configs): 65
- Source corpora used: 28
Source distribution
| OPUS corpus | rows |
|---|---|
| NLLB | 254,712,437 |
| HPLT | 23,684,089 |
| UNPC | 16,636,784 |
| CCAligned | 15,684,505 |
| MultiUN | 8,232,747 |
| MultiCCAligned | 1,274,830 |
| Tatoeba | 530,368 |
| bible-uedin | 477,291 |
| XLEnt | 457,594 |
| TedTalks | 244,550 |
| FFR | 242,670 |
| Tanzil | 195,023 |
| memat | 150,360 |
| wikimedia | 124,876 |
| KDE4 | 106,414 |
| tico-19 | 74,039 |
| MultiParaCrawl | 62,286 |
| SPC | 52,064 |
| XhosaNavy | 49,692 |
| GNOME | 45,657 |
| OpenSubtitles | 42,426 |
| infopankki | 27,305 |
| TED2020 | 18,772 |
| ParaCrawl | 14,879 |
| KDEdoc | 4,435 |
| GlobalVoices | 1,838 |
| Wikipedia | 403 |
| QED | 356 |
Language-pair distribution
| Language pair (config) | rows |
|---|---|
| english-arabic | 105,403,549 |
| english-afrikaans | 45,120,326 |
| english-swahili | 25,307,091 |
| english-amharic | 16,289,529 |
| english-somali | 10,782,645 |
| french-afrikaans | 10,481,189 |
| english-kinyarwanda | 10,022,528 |
| english-xhosa | 9,566,719 |
| english-shona | 9,471,156 |
| english-tswana | 6,274,863 |
| french-swahili | 6,006,269 |
| english-lingala | 5,823,747 |
| french-kinyarwanda | 5,325,296 |
| english-kikongo | 4,456,676 |
| english-kikuyu | 4,401,152 |
| french-somali | 3,186,029 |
| english-northern sotho | 2,900,306 |
| english-tigrinya | 2,809,183 |
| french-yoruba | 2,729,803 |
| french-chewa | 2,567,722 |
| english-tshiluba | 2,444,402 |
| french-shona | 2,338,908 |
| french-xhosa | 2,295,913 |
| french-southern sotho | 2,291,881 |
| english-mossi | 2,197,368 |
| english-fon | 1,915,450 |
| english-chewa | 1,661,884 |
| english-wolof | 1,564,588 |
| french-tswana | 1,432,179 |
| english-southern sotho | 1,379,092 |
| french-rundi | 1,267,639 |
| french-fon | 1,251,732 |
| french-tigrinya | 1,104,334 |
| french-ewe | 1,046,922 |
| english-twi | 913,853 |
| english-bemba | 854,186 |
| french-luganda | 807,504 |
| english-rundi | 776,705 |
| french-wolof | 761,196 |
| french-lingala | 708,569 |
| french-bemba | 700,722 |
| english-tsonga | 685,309 |
| french-bambara | 614,789 |
| french-kikongo | 500,755 |
| french-umbundu | 472,375 |
| french-kikuyu | 456,069 |
| english-bambara | 361,824 |
| english-berber | 353,489 |
| french-twi | 309,790 |
| english-umbundu | 298,077 |
| english-swati | 222,318 |
| english-luganda | 118,605 |
| french-berber | 53,945 |
| english-hausa | 25,338 |
| english-ewe | 15,960 |
| english-zulu | 3,699 |
| english-oromo | 3,376 |
| english-fula | 3,071 |
| english-kanuri | 3,068 |
| french-arabic | 2,537 |
| english-venda | 1,555 |
| english-igbo | 610 |
| english-malagasy | 610 |
| english-yoruba | 610 |
| french-haitian creole | 106 |
Approved corpus whitelist
OpenSubtitles, NLLB, CCMatrix, ParaCrawl, CCAligned, MultiParaCrawl, MultiHPLT, MultiCCAligned, HPLT, GNOME, LinguaTools-WikiTitles, DGT, XLEnt, WikiMatrix, UNPC, EUbookshop, ParaCrawl-Bonus, EMEA, translatewiki, MultiUN, EuroPat, KDE4, Europarl, JRC-Acquis, TildeMODEL, QED, TED2020, Samanantar, Mozilla-I10n, bible-uedin, MaCoCu, NeuLab-TedTalks, JParaCrawl, MultiMaCoCu, wikimedia, giga-fren, GoURMET, StanfordNLP-NMT, Tanzil, Anuvaad, ECB, Wikipedia, ELITR-ECA, SETIMES, DOGC, WikiTitles, Tatoeba, MBS, GlobalVoices, Finlex, News-Commentary, SciELO, PHP, JESC, ParIce, fiskmo, EOPC, MDN_Web_Docs, TED2013, EhuHac, IITB, Nunavut_Hansard, infopankki, ChuBiCo, SCB_MT_EN_TH, CAPES, MIZAN, OpenOffice, EUconst, Books, SUMMA, Elhuyar, EiTB-ParCC, TEP, ALT, Joshua-IPC, KFTT, tldr-pages, KDEdoc, WMT-News, tico-19, pmindia, ECDC, memat, hrenWaC, TedTalks, FFR, SPC, MontenegrinSubs, OfisPublik, Bianet, XhosaNavy, WikiSource, sardware, Salome, ada83, InterdialectCorpus, RF, Ubuntu, DocHPLT, liv4ever, komi
Usage
from datasets import load_dataset
# pick a language pair (config):
ds = load_dataset("African-Languages-Lab/multi-open", "english-afrikaans", split="train")
print(ds[0])
Provenance & licensing
Every row keeps its original dataset_name. Licenses vary by original source corpus — this
subset does not relicense the data; users must respect each original OPUS/source corpus license.
Ethical use
For research and educational use in advancing African-language NLP. Please use it responsibly, respect the source licenses and the communities the languages belong to.
Citation
@misc{africanlanguageslab_multiopen,
title = {African Languages Lab Multi-Open},
author = {African Languages Lab},
year = {2026},
url = {https://huggingface.co/datasets/African-Languages-Lab/multi-open}
}
- Downloads last month
- 8