Hugging Face's logo

Datasets:
goaicorp
/
DioulaMTBench

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

Request access to DioulaMTBench

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

DioulaMTBench is an evaluation set for Dioula machine translation: the FLORES+ Dioula dev and devtest sets as released, under CC BY-SA 4.0. 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.

DioulaMTBench

The established test set for Dioula machine translation, as released.
Language: Dioula, French, English 1 test set Evaluation only License: CC BY-SA 4.0

DioulaMTBench holds the FLORES+ Dioula (dyu_Latn) dev and devtest sets, joined with their English and French sides by the official FLORES+ id and copied exactly as released, so anyone can check scores against the source. Use dev for validation and devtest for final scores.

Evaluation only. Do not train or fine-tune on this dataset (FLORES+ asks the same).

Résumé en français. Le jeu de test reconnu pour la traduction automatique du dioula : les ensembles dev et devtest de FLORES+ en dioula (dyu_Latn), alignés avec leurs versions anglaise et française par l'identifiant officiel FLORES+ et copiés tels que publiés, pour que chacun puisse vérifier les scores à la source. Utilisez dev pour la validation et devtest pour les scores finaux. 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. Licence CC BY-SA 4.0, accès sur demande.

Overview

Config Source Splits Sentences Directions
flores (default) FLORES+, Open Language Data Initiative (OLDI) dev, devtest 997, 1,012 dyu ↔ fra, dyu ↔ eng

Quick start 

from datasets import load_dataset

dev = load_dataset("goaicorp/DioulaMTBench", "flores", split="dev")
test = load_dataset("goaicorp/DioulaMTBench", "flores", split="devtest")

Log in first (hf auth login) with an account that has been granted access.

How to score

  1. Apply Unicode NFC to references and outputs; change nothing else. References are stored as released.
  2. Report chrF++ (main metric, with a 95% paired bootstrap interval), BLEU and spBLEU, with the sacreBLEU signature. Use COMET only for output in French or English.
  3. Report each direction separately. Choose checkpoints on dev only; run devtest once per release candidate.

Columns

Column Description
id Official FLORES+ sentence id
eng, fra, dyu English, French and Dioula text, as released
url, domain, topic Source article and its domain (Wikinews, Wikivoyage, Wikibooks) and topic
has_image, has_hyperlink FLORES+ flags for the source sentence

Sources and credits

Config Source Version Terms
flores FLORES+, Open Language Data Initiative (OLDI) openlanguagedata/flores_plus revision 1147e2dc, dev and devtest, eng_Latn / fra_Latn / dyu_Latn CC BY-SA 4.0, not for training

See the FLORES+ dataset card for the translators and the history of the Dioula set.

How it was built

  1. Copied, not changed. Every text is identical to the FLORES+ files named above. Rows were joined across languages by their official id.
  2. Kept out of GO AI Corp's training data. No text of 4 words or more in dyu-fr-parallel or dyu-en-parallel matches a benchmark sentence after normalization (checked on 5 November 2026, exact match).
  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 set. The canary in CANARY.txt is also in every Parquet file's metadata.

Limits

  • Public set. Models trained on web data, LLMs in particular, may have seen it; their scores may be inflated.
  • English-original. Every Dioula sentence is a translation of English, so dyu→fra and dyu→eng scores translate a translation.
  • Narrow domains. Wikinews, Wikivoyage and Wikibooks topics only.
  • One spelling. Dioula orthography varies; the references follow one translator team's choices.
  • Metrics. How well chrF++ follows human judgment for Dioula is unknown.

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. The set is also available from its source above.

License

  • FLORES+ texts: CC BY-SA 4.0, as released by OLDI. Credit FLORES+ and share adaptations under the same license.
  • Our work (blocklist, metadata): CC BY-SA 4.0.

Versions

  • v1.0: FLORES+ dev and devtest, revision 1147e2dc.

Citation 

@misc{goaicorp2026dioulamtbench,
  title        = {DioulaMTBench: The established test set for Dioula 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/DioulaMTBench}
}

Please also cite FLORES+ (see its dataset card).

Contacts

Downloads last month
211
Total file size:
880 kB