Hugging Face's logo

Datasets:
goaicorp
/
MooreMTBench

Tasks:
Translation
Modalities:
Text
Formats:
parquet
Languages:
Mossi
French
English
Size:
1K - 10K
Tags:
mooré
burkina-faso
low-resource
machine-translation
benchmark
evaluation
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community

Request access to MooreMTBench

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

MooreMTBench is an evaluation set for Mooré machine translation: MAFAND, FLORES+ and BOUQuET as released, for evaluation only. The texts belong to their publishers (listed on the card) and are shared for non-commercial research only. Please read the card, its terms and licenses before requesting access, and make sure you understand them.

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

MooreMTBench

The established test sets for Mooré machine translation, in one place.
Language: Mooré, French, English 3 test sets Evaluation only GO AI Corp

MooreMTBench holds only established public test sets for Mooré machine translation, copied exactly as released so anyone can check scores against the sources, with a pinned version, a shared normalizer and a scoring recipe. Use dev splits for validation and test splits for final scores.

Evaluation only. Do not train or fine-tune on this dataset.

Résumé en français. Les jeux de test reconnus pour la traduction automatique mooré ↔ français et mooré ↔ anglais (MAFAND, FLORES+, BOUQuET), réunis au même endroit et copiés tels que publiés, pour que chacun puisse vérifier les scores à la source. Le jeu fournit aussi une version figée, un normaliseur et une méthode de calcul des scores communs. Aucune de ses phrases n'est dans les données d'entraînement de GO AI Corp. Réservé à l'évaluation : n'entraînez et n'affinez aucun modèle sur ces données. Accès sur demande.

Overview

Config Content Written first in Best for Validation Test
mafand-fr-mos (default) Burkinabè news (Burkina24, Lefaso), MAFAND-MT French fra→mos 1,492 1,574
flores Wikinews, Wikijunior, Wikivoyage, FLORES+; eng / fra / mos English eng→mos; comparison with published systems 997 (dev) 1,012 (devtest)
bouquet Conversation, how-to, narration, social posts, comments, web, reflection, BOUQuET; eng / fra / mos Other languages (e.g. Egyptian Arabic) Everyday registers; comparison with BurkimbIA's leaderboard 504 854

A set is source-original only in the direction its source side was written in first; in the others the model translates a translation, which usually scores higher than real use. Report each score with its direction.

Quick start 

from datasets import load_dataset

test = load_dataset("goaicorp/MooreMTBench", "mafand-fr-mos", split="test", revision="v1.0")

Always pin revision so scores stay comparable.

How to score

  1. Normalize references and outputs with tools/moore_norm.py (moore-norm 1.0): Unicode NFC, Greek look-alikes ε ι υ mapped to ɛ ɩ ʋ, tildes, apostrophes, hyphens and spaces unified, escaped quotes (') removed, nothing else changed. References are stored as released (MAFAND and FLORES+ Mooré write most tildes as separate combining marks, MAFAND French uses curly apostrophes), so skipping this step lowers scores.
  2. Score with sacreBLEU: chrF++ first (chrF2++, word order 2), BLEU second (13a). Report the signature.
  3. Compare two systems with paired bootstrap resampling (1,000 samples). Differences of 1-2 points are often noise.
  4. Report each config, split and direction separately, plus one overall score per direction: a single corpus-level score over the three test splits pooled (3,440 sentences: 1,574 + 1,012 + 854). Do not average the per-config scores, which would weight a BOUQuET sentence almost twice as much as a MAFAND one.

Cite a score like this: MooreMTBench v1.0 · mafand-fr-mos test · fra→mos · chrF2++ <signature> · moore-norm 1.0. For the overall score: MooreMTBench v1.0 · overall test (pooled) · fra→mos · chrF2++ <signature> · moore-norm 1.0.

Columns

Config Columns
mafand-fr-mos id, fr, mos, mafand_index (line number, from 0, in MAFAND's official file)
flores id, eng, fra, mos, url, domain, topic, has_image, has_hyperlink
bouquet id, par_id (paragraph), domain, register, eng, fra, mos, orig_text (the sentence in the language it was written in)

Sources and credits

Config Source Exact version Terms
mafand-fr-mos MAFAND-MT (Adelani et al., NAACL 2022), Masakhane masakhane-io/lafand-mt commit 9b698bb, data/json_files/fr-mos/dev.json and test.json CC BY-NC 4.0
flores FLORES+, Open Language Data Initiative (OLDI) openlanguagedata/flores_plus revision 1147e2dc, dev and devtest, eng_Latn / fra_Latn / mos_Latn CC BY-SA 4.0, not for training
bouquet BOUQuET (Omnilingual MT Team, Meta, 2025) facebook/bouquet revision dac03947, sentence level, dev (here validation) and test CC BY 4.0

How it was built

  1. Copied, not changed. Every text is byte-identical to the file named above. Rows were only joined across languages by their official ids (FLORES+ id, BOUQuET uniq_id) and given an id.
  2. Kept out of GO AI Corp's training data. Every row of MooreFrenchParallel, mos-fr-parallel-internal and MooreEnglishParallel was removed whose French, English or Mooré equals a benchmark sentence (4 words or more) or shares at least half of its 6-word sequences with one (9 words or more), and is listed in that dataset's decontamination/ folder. Short common phrases ("Bonjour", "Oui, exactement") were kept after a manual check. Among others, this removed 1,012 FLORES+ devtest pairs left in an older collection.
  3. Blocklist and canary. blocklist.sha1.txt holds a hash of every source and reference, so training pipelines can drop matches without reading the test sets. The canary in CANARY.txt is also in every Parquet file's metadata.

Limits

  • All three sets are public. Models trained on web data, LLMs in particular, may have seen them; their scores may be inflated.
  • No Mooré-original test. MAFAND is French-original, FLORES+ English-original and BOUQuET written in other languages, so every mos→fra and mos→eng score translates a translation.
  • Narrow domains. News, Wikipedia-style topics and BOUQuET's everyday registers. No test for health, administration or religion.
  • Translation quality. No native-speaker audit of the FLORES+ or BOUQuET Mooré references has been published.
  • Metrics. How well chrF++ follows human judgment for Mooré is unknown.

Planned: a Mooré-original test, written and reviewed for this purpose and released openly.

Access

Fill in the form on this page. We approve people and institutions that say who they are, describe an evaluation use and accept the terms. Each set is also available from its source above.

License

  • Evaluation only: no training or fine-tuning on any part of this dataset.
  • Texts: each config keeps its source's license (see Sources and credits).
  • Our work (normalizer, blocklist, metadata): CC BY-NC 4.0.

Versions

Tags follow vMAJOR.MINOR. A minor version adds a config or a tool. A major version changes rows of an existing config, and its scores cannot be compared with earlier ones. Released files are never edited without a new tag.

  • v1.0 (November 2026): MAFAND, FLORES+, BOUQuET.

Citation 

@misc{goaicorp2026mooremtbench,
  title        = {MooreMTBench: The established test sets for Mooré machine translation},
  author       = {Nyantudre, Wendyellé Abubakrh Alban and Bandaogo, Aristide},
  year         = {2026},
  howpublished = {Hugging Face Datasets (gated access), GO AI Corp},
  url          = {https://huggingface.co/datasets/goaicorp/MooreMTBench}
}

Please also cite the source of each set you use: MAFAND-MT (Adelani et al., 2022), FLORES+ (NLLB Team et al., 2022; OLDI) and BOUQuET (Omnilingual MT Team, 2025).

Contacts

Downloads last month
308
Total file size:
2.15 MB