burkimbia/moore-parallel-corpus

Mooré–French parallel corpus

French–Mooré sentence and term pairs for machine translation, built by bia-datasets-text from 18 cooked sources: cleaned, orthography-normalized, annotated, filtered and deduplicated, with the frozen BurkimbIA MT benchmark's source texts excluded from every split.

from datasets import load_dataset

ds = load_dataset("burkimbia/moore-parallel-corpus", revision="v0.9.0")   # 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	120,957
val	1,180
test	2,496

segment_type	train	val	test
lexical-gloss	29,365	0	0
definition	131	1	1
sentence	65,159	838	1,782
passage	25,067	331	690
turn	1,235	10	23

domain	train	val	test
religious	72,284	946	2,028
news	5,023	84	135
government	2,785	20	51
legal	377	2	3
health	1,461	8	13
technology	666	7	13
agriculture-food	913	0	1
education	1,171	2	13
oral-literature	1,867	16	58
everyday	3,395	27	49
general	31,015	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.

Columns

column	meaning
french	French text.
moore	Mooré text.
source	Where the pair comes from; families below.
source_lang / target_lang	Always fr → 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 for each side; und for numbers and punctuation, null for lexical-gloss pairs.

Sources

family	train	val	test	origin	terms
Bible (bible.com)	36,251	528	1,128	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)	29,366	336	724	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	14,935	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.
French–Mooré index	9,768	0	0	French–Mooré dictionary index (PDF), read from its text layer.	Rights held by the publisher.
Quran	6,597	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,	Rights held by the translators and publishers.
MAFAND-MT	5,023	84	135	Masakhane's MAFAND-MT news translations, French–Mooré, via the MooreFRCollections mirror.	CC BY-NC 4.0 (non-commercial).
SMOL	5,035	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.
Council of Ministers	2,718	19	49	Council of Ministers communiqués, Government of Burkina Faso (sig.gov.bf), French and Mooré editions.	Official government publication.
moore-web reviewed pairs	2,187	16	35	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.
Expert translations	2,106	13	22	Translations commissioned by BurkimbIA from professional translators (legal, health, software manuals, radio, children's rights).	BurkimbIA.
Bilingual books and lexicons	2,027	1	5	Bilingual books and thematic lexicons (health, agriculture and livestock, food, animals, body parts, a reading book), transcribed to spreadsheets.	Rights held by the publishers.
Montivilliers–Nassere dictionary	1,242	0	0	Online dictionary of montivilliersnassere.fr, via the MooreFRCollections mirror.	Website terms.
SPG series	1,164	9	22	Interviews and tales from the SPG TV series (YouTube), transcribed and translated.	BurkimbIA translations of third-party broadcasts.
Mooré proverbs	892	9	41	Mooré proverbs with French translations, media.ipsapps.org.	Rights held by the publisher.
Surveys	566	12	23	Survey answers collected and translated by BurkimbIA.	BurkimbIA.
Seed sentences	493	6	6	Translator-checked seed sentences and dialogue turns.	BurkimbIA.
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	238	0	0	Greetings curated by a Mooré speaker; numbers spelled by the pipeline's Mooré number speller.	BurkimbIA.

The frozen MT benchmark of BurkimbIA is not part of this dataset; its source texts are excluded from every split.

Changes in v0.9.0

Datasets built from this release

• The train rebuilt with main.py --lid (S3 version 20261010T144128Z-886bd11) and burkimbia/moore-parallel-corpus tagged v0.9.0: the same 124 633 pairs as v0.8.1, 0 pairs change split (train 120 957, val 1 180, test 2 496). New columns: domain, and the four GlotLID columns, null on the 29 365 lexical-gloss pairs. On the 95 268 other pairs GlotLID labels 99.1 % of French sides fra_Latn and 98.9 % of Mooré sides mos_Latn
• No LASER or COMET-QE scores: on the available RTX 3050 (4 GB), scoring the 95 268 non-gloss pairs measured ~3.7 h for LASER and ~8.5 h for COMET-QE. They will come in a later release scored on a larger GPU; LaBSE as a LASER replacement is #113

Fixed

• Run COMET-QE prediction once over the full dataset after batched LASER scoring (#47), rather than once per LASER map batch. COMET keeps its own inference batch size, score rounding and post-scoring thresholds; empty inputs skip inference and re-scoring replaces existing score columns. lexical-gloss pairs are no longer scored: they get null scores and pass both thresholds, since sentence embeddings and COMET-QE say nothing reliable about a lone word. Score columns are always float64, empty datasets included. With --gpus above 1, COMET's DDP padding is trimmed and any other length or ordering mismatch raises instead of misaligning scores. comet_qe.score_pairs_qe shares the same prediction helper and alignment check; the score thresholds filter in batches

Added

• The publisher carries laser_score, comet_qe and the four GlotLID columns when main.py ran with --score / --lid, and the dataset card describes them only then. GlotLID now leaves lexical-gloss pairs null, like LASER and COMET-QE: a lone word is too short for a reliable prediction
• domain column on train, rejected and published pairs: what the pair is about (religious, news, government, legal, health, technology, agriculture-food, education, oral-literature, everyday, general), read from the source value with ordered prefixes in config/domains.yaml. Independent of segment_type; mixed sources (dictionaries, SMOL, reviewed pairs) are general. main.py and publish_annotated_split.py stop on a source value no prefix covers. The publisher recomputes it after recovering v1 provenance, and the dataset card gains a domain × split table. No re-cook needed. On the 2026-10-10 train: 74 853 religious, 31 271 general, 5 150 news, 4 497 everyday, 2 840 government, 1 926 oral-literature, 1 379 health, 1 067 education, 686 technology, 585 agriculture-food, 379 legal
• Optional GlotLID v3 language identification during consolidation (#48). main.py --lid predicts the language/script and confidence of both sides in batches, retaining source_lid_label, source_lid_score, target_lid_label and target_lid_score in train and rejected outputs. The model is downloaded and cached from cis-lmu/glotlid (model_v3.bin), or loaded from --lid-model; --lid-batch-size controls inference batches. Non-linguistic inputs receive und with zero confidence. Predictions annotate pairs without changing their declared languages or introducing automatic language rejection (ADR-008, ADR-015)
• Document GlotLID setup, local-model usage, output columns and the successful full local run on 2026-10-10: 146 677 input rows from 18 sources, 124 633 train pairs and 5 523 rejected candidates

Known issues

• The SMOL French side is a translation of the English, not an original French text (#99).
• Language codes are fr and mos; a move to ISO 639-3 fra is planned (#100).

Provenance

Release	v0.9.0 of bia-datasets-text (08d2a92)
S3 train version	20261010T144128Z-886bd11
S3 split version	20261010T145350Z-08d2a92

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