You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Mooré–French parallel corpus

French–Mooré sentence and term pairs for machine translation, built by bia-datasets-text from 20 cooked sources: cleaned, orthography-normalized, annotated, filtered and deduplicated.

from datasets import load_dataset

ds = load_dataset("burkimbia/moore-parallel-corpus", revision="v0.13.0")   # this release; omit revision for the latest

Usage and license

Research and non-commercial use only. The corpus aggregates sources with different rights: each keeps its own license or terms, listed per source below. Some are licensed for non-commercial use only (MAFAND-MT), and several are copyrighted works included for research under gated access, not under an open license. Access is granted manually; by requesting it you agree to respect each source's terms.

Splits

split pairs
train 128,470
val 1,221
test 2,601
segment_type train val test
lexical-gloss 29,814 0 0
definition 131 1 1
sentence 71,333 866 1,868
passage 25,956 344 709
turn 1,236 10 23
domain train val test
religious 72,591 946 2,032
news 2,331 28 51
government 11,450 114 224
legal 426 2 3
health 1,708 9 15
technology 778 7 14
agriculture-food 755 0 1
education 1,149 2 13
oral-literature 2,010 17 60
everyday 4,744 28 56
general 30,528 68 132

How pairs are assigned:

  • Pairs held out by the first published version stay held out (anchor table, matched on a tolerant key).
  • Every other pair is placed by a hash of its normalized text, 1 % val and 2 % test, so the same pair always lands in the same split, and a punctuation, spacing or case correction does not move it (ADR-016).
  • All translations of a French entry go to the same split, and train wins (ADR-007).
  • lexical-gloss pairs (dictionary headwords, glossary terms, numbers, greetings) are never held out: they test vocabulary lookup, not translation.
  • A pair with a near copy in train, on either side, goes to train.
  • A pair in the previous release's train stays in train: a model trained on that release has seen it.

This repository replaces burkimbia/fr_mos_annotated_split_v2 to _v4. Its val/test are not comparable with those: since ADR-016 a pair's split is hashed from its normalized text, which redrew the pairs the v1 anchor does not pin (5 164 pairs moved against v4). Within this repository, compare test scores only across PATCH and MINOR releases, except before 1.0.0: a 0.x MINOR may change the test set and says so in its changes (v0.10.0 does).

Columns

column meaning
src Source text, French: the side source_lang names.
trg Target text, Mooré: the side target_lang names.
source Where the pair comes from; families below.
source_lang / target_lang Always fra → mos: every pair is turned to French→Mooré.
original_lang The language the source wrote first: mos for pairs a dictionary or a tale wrote Mooré→French.
segment_type lexical-gloss, definition, sentence, passage or turn.
domain What the pair is about, from its source: religious, news, government, legal, health, technology, agriculture-food, education, oral-literature, everyday, general.
id / original_id Row identifiers.
was_corrected / quality_warnings Human-review signal inherited from v1, when the pair existed there.
len_ratio Length ratio of the two sides, as used by the train filters.
source_lid_label / source_lid_score / target_lid_label / target_lid_score GlotLID v3 top language/script and confidence (0 to 1, 4 decimals) for each side; und for numbers and punctuation, null for lexical-gloss pairs.

Sources

family train val test origin terms
Bible (bible.com) 36,386 527 1,130 Catholic and Protestant Mooré Bible versions from bible.com, aligned with French by book, chapter and verse. Rights held by the translations' publishers.
Bible (New World Translation) 29,494 337 726 Bible verses in French and Mooré from jw.org, via the sawadogosalif/MooreFRCollections mirror. jw.org terms of use. The publisher grants no redistribution license; the mirror's MIT label does not cover this text.
Niggli dictionary 15,336 34 70 Urs Niggli, Mooré–French–English dictionary (2021). Headwords and example sentences read from the PDF text layer. Rights held by the author and publisher.
Council of Ministers 11,372 112 222 Council of Ministers communiqués, Government of Burkina Faso (sig.gov.bf), French and Mooré editions, aligned sentence by sentence by moore-web (LASER, filtered on LASER and COMET-QE) and not reviewed by a person. Official government publication.
French–Mooré index 9,816 0 0 French–Mooré dictionary index (PDF), read from its text layer. Rights held by the publisher.
Quran 6,641 79 174 Mooré and French Quran translations aligned by sura and aya; verses over 526 characters segmented by GPT-4o, segments paired by position (the cartesian pairing fixed in v0.8.1, #98). Rights held by the translators and publishers.
SMOL 4,971 33 61 Google SMOL. The Mooré was translated from English; the French side is a translation of the English, most likely by machine, so these pairs are not French→Mooré translations (#99). CC BY 4.0.
Expert translations 3,324 14 23 Translations commissioned by BurkimbIA from professional translators (legal, health, software manuals, radio, children's rights). BurkimbIA.
moore-web reviewed pairs 2,847 17 39 Pairs reviewed and corrected by a human in moore-web's review app (SIDA book and its facilitator guide, reading book, UDHR, tales, New Year messages). Rights held by each document's publisher.
MAFAND-MT 2,331 28 51 Masakhane's MAFAND-MT news translations, French–Mooré, from the upstream files (burkimbia/mafand-fr). CC BY-NC 4.0 (non-commercial).
Bilingual books and lexicons 1,440 1 3 Bilingual books and thematic lexicons (health, agriculture and livestock, food, animals, body parts, a reading book), transcribed to spreadsheets. Rights held by the publishers.
SPG series 1,165 9 22 Interviews and tales from the SPG TV series (YouTube), transcribed and translated. BurkimbIA translations of third-party broadcasts.
Mooré proverbs 1,035 10 43 Mooré proverbs with French translations, media.ipsapps.org. Rights held by the publisher.
Surveys 853 13 30 Survey answers collected and translated by BurkimbIA. BurkimbIA.
Seed sentences 493 6 6 Translator-checked seed sentences and dialogue turns. BurkimbIA.
Montivilliers–Nassere dictionary 366 0 0 French–Mooré lexicon of montivilliersnassere.fr, read from the text layer of its PDF (2012, archived copy). Website terms.
Digital and postal glossary 349 1 1 Ministry of Digital Affairs and Post, glossary of digital and postal terms in Mooré, terms and definitions aligned with the French lexicon. Official government publication.
Greetings and numbers 251 0 0 Greetings curated by a Mooré speaker; numbers spelled by the pipeline's Mooré number speller. BurkimbIA.

Changes in v0.13.0

Datasets built from this release

  • The train rebuilt with main.py --lid from 8809ec6, the code of this release before the version bump (train version 20261105T134100Z-8809ec6): 132 292 pairs and 5 829 rejected, as v0.12.0. burkimbia/moore-parallel-corpus tagged v0.13.0: train 128 470, val 1 221, test 2 601, no pair changes split against v0.12.0, same columns
  • 5 297 published pairs change source against v0.12.0: 5 281 from the publisher no longer renaming pairs to v1's name (below: 1 590 dictionnary-index → Niggli, 1 432 the other way, 1 058 montivilliers → moore_tv, 526 reading book → moore-web/du-moore, 97 ministers_council → moore-web/conseils, ...), and 16 reviewed pairs that the v1 text join had given the source of another row. 1 566 pairs change domain with them, 1 086 of them general → everyday
  • GlotLID columns as in v0.12.0, scores now at most 1 (below). Still no LASER / COMET-QE scores

Added

  • Offline MT mixture builder (python -m bia_datasets_text.mix) with strict typed recipes, attribute-based components, lexical-gloss-aware direction policies, exact quotas, single-attribute balancing, and deterministic bounded repetition. Preserves published validation/test splits and provenance; writes versioned, reusable datasets plus recipe and build manifests for bia-llms and bia-nllb.
  • directions.prefer_original for explicit mixture directions: each direction uses the pairs originally written in its source language before reversing others to complete its quota, within each balance stratum. On v0.12.0 at 1 000 examples with a 50/50 fra/mos recipe balanced on domain, mos→fra went from 33 Mooré originals to 62 of 425: Mooré originals exist only in general, oral-literature and one education pair, so the other domains' mos→fra quotas are still filled by reversal.
  • AGENTS.md gains the branch, commit, CHANGELOG, label and PR conventions shared with the other BurkimbIA repos, and TESTING.md says what a test should check and where to fake S3, the Hub and the scoring models.

Changed

  • LASER scoring uses nano-laser (BurkimbIA's plain-torch port, same API and checkpoints) instead of Meta's laser-encoders, which pulled in fairseq. Same embeddings on French and Mooré sample pairs (max difference 1.2e-7) and same LASER scores. The scoring extra moves to torch>=2.6,<2.11 (2.10.0, CUDA 12.8; 2.11 is the first PyPI build on CUDA 13) and pins setuptools<82, the last with pkg_resources, which COMET's torchmetrics imports. COMET-QE (McGill-NLP/ssa-comet-qe) gives the same scores on torch 2.10 as on 2.5.1
  • Python 3.12 or newer (requires-python, .python-version), as in the Dockerfile and CI
  • Move publish_annotated_split.py to scripts/: run it as uv run python scripts/publish_annotated_split.py [--push]. Tests import it through pythonpath = ["scripts"] in pyproject.toml. No change to the published dataset
  • Move phrases_condensees_a_revoir.csv to data/, which now also holds review worklists the code does not read
  • labelstudio_reviewed reads v1 (burkimbia/fr_mos_annotated_split) once, at a pinned revision, and keeps it on S3 (raw: cache) as two tables: the 5 718 reviewed pairs and was_corrected / quality_warnings for every v1 row. The reviewed pairs keep v1's source from cooking on, so cleaned_data carries their real provenance and domain instead of labelstudio_reviewed / general (3 770 train rows). The publisher no longer recovers source or recomputes domain, reads v1 from that S3 copy instead of the Hub, and refuses a train still holding the internal tag. Its text join had given 16 reviewed pairs the source of a different, unreviewed v1 row (algue → toe-neega as dictionnary-index instead of livre- Nourriture); 2 of them change domain. Train rows, splits, was_corrected and quality_warnings are unchanged. Needs cook labelstudio_reviewed then main.py before the next publication

Fixed

  • MT mixing accepts case/punctuation variants within a split and the example recipe supports every published language pair. Preserve Arrow feature types, clear reversed COMET-QE scores before score filtering/balancing, treat non-finite scores as missing, and compare numeric membership values numerically. Compile filters once and stream corpus fingerprints during validation. Recipe loading uses the central configuration boundary and fingerprints nested/external recipes; versioned outputs keep one authoritative manifest and remove obsolete stable metadata copies.
  • The publisher no longer renames a pair's source to the name v1 gave it. build_dataframe took v1's source for every pair it found in v1, where only labelstudio_reviewed pairs need it, so on the current train 5 281 pairs went out under another name, often a retired one: 97 moore-web/conseils as ministers_council, 11 moore-web/sida as the poko book, 526 moore-web/du-moore as the reading book, 1 058 Expert: manual_alignement.xlsx::moore_tv as dictionnary-swww.montivilliersnassere.fr, about 3 000 swapped between Niggli and the dictionary index. Most are the same pair in two sources: the publisher replaced the copy main.py keeps (trusted first), and with it the pair's domain and card family, by the copy v1 had kept. was_corrected and quality_warnings still come from v1. Text and splits are unchanged. Takes effect at the next published version
  • GlotLID scores no longer exceed 1. fastText's predict() returns exp(log(p + 1e-5)), so every probability came back 1e-5 too high: 84 307 Mooré-side and 3 214 French-side scores in v0.12.0 were above 1, up to 1.00001. language_id now removes the offset, clips to [0, 1] and keeps 4 decimals (FASTTEXT_OFFSET, SCORE_DECIMALS). Scores only; labels and kept pairs are unchanged. Takes effect at the next main.py --lid run and published version
  • The card's known issues no longer say the language codes are fr and mos with a move to ISO 639-3 planned (#100): the published values are fra and mos since v0.10.0. The line is on the v0.10.0 and v0.12.0 cards

Documentation

  • Bible JW logbook: Salif confirms its verses were never aligned, only scraped side by side. SMOL logbook: whether to keep its machine-translated French pivot is left to training experiments. Both stay in moorefr_collection, tracked on #128; #99 is closed

Security

  • Bumped certifi to 2026.10.19 in uv.lock and the uv-pre-commit hook to uv 0.13.4. Earlier versions ship the compromised emSign ECC Root CA - G3 root, which certifi 2026.10.19 and uv 0.13.4 remove.

Known issues

  • The SMOL French side is a translation of the English, not an original French text (#99).

Provenance

Release v0.13.0 of bia-datasets-text (f36b8a3)
S3 train version 20261105T134100Z-8809ec6
S3 split version 20261105T140702Z-f36b8a3

Built by bia-datasets-text at the tag above. Contact: BurkimbIA, https://github.com/BurkimbIA/bia-datasets-text/issues.

Downloads last month
1,154