burkimbia/mt-benchmark-public

Burkina languages MT leaderboard evaluation set

The sets the Burkina languages MT leaderboard scores translations on: French↔Mooré from BOUQuET, French↔Dioula from FLORES+. The leaderboard is hosted by BurkimbIA and open to every team, in Burkina Faso or elsewhere. The rows are those of the originals, unchanged; only the columns are renamed.

One config per benchmark and language. BOUQuET and FLORES+ give a sentence the same id in every language, so a direction is two configs of the same benchmark joined on id, and adding a language is adding a config.

from datasets import load_dataset

fra = load_dataset("burkimbia/mt-benchmark-public", "bouquet_fra", split="test", revision="v1.0.0").to_pandas()
mos = load_dataset("burkimbia/mt-benchmark-public", "bouquet_mos", split="test", revision="v1.0.0").to_pandas()
pairs = fra.merge(mos, on=["id", "domain"], suffixes=("_fra", "_mos"))  # French→Mooré, or back

To submit a model, translate every text of the source language, one translation per id, and send the translations on the leaderboard.

Usage and license

Evaluation of machine translation. Each config keeps the license of the benchmark it comes from: the bouquet_* configs are BOUQuET's (Meta), under CC BY 4.0; the flores_plus_* configs are FLORES+'s (Open Language Data Initiative), under CC BY-SA 4.0, which any copy or derivative of them keeps. Access is granted automatically, as for the originals, to keep the test sets out of web crawls; by requesting it you agree to respect each license. These are test sets: a model trained on them has meaningless scores.

Configs

config	language	from	rows
bouquet_fra	French	BOUQuET, sentence level, test	854
bouquet_mos	Mooré	BOUQuET, sentence level, test	854
flores_plus_fra	French	FLORES+, devtest	1,012
flores_plus_dyu	Dioula	FLORES+, devtest	1,012

Directions

direction	source	reference
fra→mos	bouquet_fra	bouquet_mos
mos→fra	bouquet_mos	bouquet_fra
fra→dyu	flores_plus_fra	flores_plus_dyu
dyu→fra	flores_plus_dyu	flores_plus_fra

Why these sets

Teams that build translation models, BurkimbIA among them, usually evaluate them on their own data, drawn from the sources they train on. Scores on such data reward contamination and a shared distribution as much as translation quality, and cannot rank models from different teams. The leaderboard scores every model on sets meant to be independent of everyone's training data, BOUQuET and FLORES+, assuming that no team trains on them. BOUQuET has no Dioula, hence FLORES+ for it. Scores do not compare across languages: the two benchmarks differ.

Columns

column	meaning
id	BOUQuET's uniq_id (paragraph and sentence, e.g. P002-S1), or FLORES+'s id: the same in every config of a benchmark.
text	The sentence, in the config's language.
domain	The benchmark's own domain: eight in BOUQuET, the source wiki in FLORES+.

BOUQuET:

domain	rows
conversation	186
how-to | instructions	107
social posts	104
web misc.	103
narration	98
other misc.	95
reflection	88
comments	73

FLORES+:

domain	rows
wikibooks	351
wikinews	341
wikivoyage	320

Citation

• bouquet_*: cite BOUQuET, Andrews et al., EMNLP 2025.
• flores_plus_*: cite FLORES-200 and the papers of FLORES+'s bibliography.

Provenance

Version	v1.0.0
BOUQuET	as of 2026-10-14
FLORES+	version 4.6

Contact: contact@burkimbia.com