MooreFrenchParallel collects publicly available French-Mooré parallel text for machine translation research. The texts belong to their publishers (listed on the card) and are shared for non-commercial research only; our work is CC BY-NC 4.0. Please read the card, its terms and licenses before requesting access, and make sure you understand them.

MooreFrenchParallel

Publicly available French-Mooré parallel text for machine translation.

MooreFrenchParallel gathers the publicly available French-Mooré parallel text we know of: news, Bibles, the Quran, government communiqués, dictionary examples, proverbs, word lists, the UDHR and text mined from the web.

Access is gated and requests are reviewed by hand. Collecting and cleaning Mooré parallel text takes a lot of time and resources, and we do it mostly out of passion. Please read this card, its terms and licenses before requesting access, and make sure you understand them. Respecting them keeps this work useful to the community and helps move our local languages forward.

Résumé en français. Textes parallèles français-mooré disponibles publiquement, pour la traduction automatique : presse, Bibles, Coran, comptes rendus du Conseil des ministres, exemples de dictionnaire, proverbes, lexiques, la Déclaration universelle des droits de l'homme et des textes collectés sur le web. Accès sur demande, examinée à la main, réservé à la recherche non commerciale. Collecter et nettoyer ces textes demande beaucoup de temps et de moyens, et nous le faisons surtout par passion : lisez cette carte, ses conditions et ses licences avant de demander l'accès.

Overview

Subset	Content	Source	Terms	Pairs
`mafand-news`	News	MAFAND-MT fr-mos (Adelani et al., NAACL 2022), Masakhane	CC BY-NC 4.0	2,520 (MAFAND train)
`bible-mpbu`	Bible, verse-aligned	Mooré: WẼNNAAM SEBRE 2014, © Bible Society of Burkina Faso (bible.com). French: Louis Segond 1910, public domain	bible.com: personal, non-commercial use	30,771
`bible-sebr-songo`	Bible with deuterocanon, verse-aligned	Mooré: SEBR SÕNGO, © Alliance Biblique du Burkina Faso, 1998. French: Louis Segond 1910 (public domain); deuterocanon néo-Crampon 2022 (CC BY-SA 4.0) and Crampon 1923 (public domain)	Mooré: all rights reserved; néo-Crampon: CC BY-SA 4.0	34,897
`bible-jw-nwt`	Bible, verse-aligned	Mooré and French New World Translation (French 2018), © Watch Tower Bible and Tract Society, from the official jw.org EPUB files	jw.org: terms of use of the site	30,663
`webonary-examples`	Dictionary example sentences	Dictionnaire Mooré - français - English avec images, compiled by Urs Niggli, © 2016 SIL International	No license stated	4,236
`smol`	Words and sentences	Google SMOL / GATITOS (Mooré side, CC BY 4.0); French from GATITOS for words, added by us for older rows, and machine-translated from the English (OPUS-MT) for the SmolSent and SmolDoc sentences added in November 2026 (all flagged)	CC BY 4.0	11,301
`udhr`	Universal Declaration of Human Rights	UN OHCHR translation, via the Unicode UDHR project	Free use	160
`razamwende`	Reading lessons	razamwende reading site	No terms found	34
`proverbs-kabore`	Proverbs	Proverbes en mooré vol. 1 to 8, French by Timothée Kaboré, published by Urs Niggli (Sagesse Africaine)	No license stated	322
`webonary-dictionary`	Dictionary words and glosses	Urs Niggli's Mooré dictionaries (Webonary 2021, © 2016 SIL International; Mooré Lexique 2017)	No license stated	34,940
`lexicon-old-spelling`	Word lists in old spelling	Probably the montivilliersnassere.fr lexicon (to confirm)	No terms found	1,298
`webonary-articles`	Raw dictionary articles (whole entry in the French field, not sentence pairs)	Urs Niggli, Mooré Lexique 2017	No license stated	2,039
`quran-quranenc`	Quran, verse-aligned	Mooré: Rowwad Translation Center. French: Rachid Maach. Both translated from the Arabic, via QuranEnc	No license stated	6,142
`gov-bf-communiques`	Council of Ministers reports, sentence-aligned	French originals and Mooré versions, © Government of Burkina Faso, Porte-parole du Gouvernement (gouvernement.gov.bf), 2024 to 2026	All rights reserved	14,377
`nllb`	Web-mined sentences	NLLB mined English-Mooré bitext (Meta, 2022, via allenai/nllb); French machine-translated from the English (OPUS-MT). Probably includes text from jw.org, as in the original NLLB release.	NLLB: ODC-By 1.0; texts belong to their websites	45,317

219,017 unique pairs in 15 subsets, one per source. Each pair appears once, under its primary source. default holds the 135,423 sentence-level training pairs of 11 subsets; dictionaries, word lists, raw dictionary articles and the web-mined nllb subset loads on its own.

Quick start

from datasets import load_dataset

ds = load_dataset("anyantudre/MooreFrenchParallel", split="train")                         # default
bible = load_dataset("anyantudre/MooreFrenchParallel", "bible-mpbu", split="train")       # one subset

Columns

Column	Description
`id`	Stable id (hash of the normalized pair)
`fr`, `mos`	French and Mooré text
`source`	Subset of the pair
`unit`	`word`, `phrase`, `sentence`, `verse` or `paragraph`
`ref`	Bible reference (e.g. `GEN.1.1`) or entry id, when known
`french_version`	French Bible of verse pairs (`lsg1910`, `nwt-fr-2018`, ...)
`fr_note`	Explanation that came with a proverb, split off the French
`in_mafand_eval`	True if the French or Mooré equals a MAFAND dev or test sentence (here only short phrases like "Bonjour")
`flags`	Quality notes, e.g. `swapped_fixed`, `ratio_extreme`, `fr_machine_translated`, `mined_bitext`, `agent_checked`, `lower_confidence`; `speechbench_heldout` marks verses of the Bible chapters held out by MooreSpeechBench
`origins`	Earlier datasets or URLs the pair came from

How it was built

Input: all French-Mooré collections we have been gathering since 2024, plus Bibles, dictionary examples, the Quran, government communiqués, SMOL sentences and NLLB text collected in November 2026.

Normalize: Unicode NFC, spacing, escaped quotes, swapped languages and PDF ligatures fixed; empty and URL-only rows dropped.
Label: each pair gets its real source, public sources first when a pair is in several.
Deduplicate: one row per pair, ignoring case and punctuation.
Remove test data: MAFAND dev and test, FLORES+ and BOUQuET are left out and published as released in MooreMTBench. Removed: rows that repeat one of their sentences (4+ words) or share at least half of their 6-word sequences with one (9+ words), 116 rows (list in decontamination/). Short common phrases ("Bonjour") were kept.
Align Bibles and the Quran verse by verse: versification differences merged, verse bridges kept as one pair.
Align the communiqués: both PDFs are cut at their shared section numbers, then sentences are linked with the help of a French-to-Mooré translation model (used only to score links, never kept). A language model re-read every weak link against both PDFs. Pairs it rejected or could not settle, sentences cut by page breaks and loose summaries were dropped.
Review: automatic checks propose drops (length ratio, identical sides, machine-translation loops); when a pair is doubtful, it is dropped.

Evaluation

Training only: every subset has one train split. Evaluate with MooreMTBench.

Limits

Bible verses are most of the data, so the vocabulary leans religious.
The Bible subsets include the chapters held out by MooreSpeechBench v1.0 (list in anyantudre/MooreSpeechBench, sets/heldout_chapters.csv), flagged speechbench_heldout; nllb may hold some of them too. Exclude them before training any speech system's text components (ASR language models, TTS text front ends).
Dictionary subsets are mostly single words; leave them out of sentence-level training if you do not want them.
Some Mooré text uses older spellings, and some French was machine-translated (flagged fr_machine_translated): all of nllb and part of smol.
nllb was mined automatically from the web; no person checked these pairs, and some are loose matches.
gov-bf-communiques is administrative text (decrees, nominations); the Mooré sometimes shortens the French.

Access

Fill in the form on this page. We approve people and institutions that say who they are, describe a research use and accept the terms. Anonymous or vague requests will not be approved.

License

Texts: property of the owners listed above, with no license granted; shared for non-commercial research, and owners can ask for removal.
MAFAND-MT: CC BY-NC 4.0. SMOL / GATITOS: CC BY 4.0. néo-Crampon French rows: CC BY-SA 4.0. NLLB bitext: ODC-By 1.0 (the texts stay with their websites).
Our work (alignment, cleaning, labels, metadata): CC BY-NC 4.0.

For other uses or questions, contact us: Wendyellé A. Alban NYANTUDRE (nyantudrealban@gmail.com). By accessing this dataset you agree to these terms.

Citation

@misc{nyantudre2026moorefrenchparallel,
  title        = {MooreFrenchParallel: Publicly available French-Mooré parallel text for machine translation},
  author       = {Nyantudre, Wendyellé Abubakrh Alban and Bandaogo, Aristide},
  year         = {2026},
  howpublished = {Hugging Face Datasets (gated access)},
  url          = {https://huggingface.co/datasets/anyantudre/MooreFrenchParallel}
}

If you use a subset, please also cite its source (see Overview).

Downloads last month: 546

Total file size:

35.3 MB