burkimbia/fr-moore-reasoning-v1

FR-Mooré Reasoning v1

Bilingual French ↔ Mooré (mos) dataset pairing each translation with a reasoning trace in French that explains, structurally, how one side maps to the other. Built by BurkimbIA to train reasoning-capable translation models for Mooré, a low-resource language of Burkina Faso.

Dataset at a glance

Records	248 766 (236 327 train / 12 439 test)
Source pairs	124 383, each expanded into both directions (fr→moore, moore→fr)
Source dataset	burkimbia/fr_mos_annotated
Reasoning language	French
Split	95% / 5%, seed=42

Fields

Field	Type	Description
input_lang	string	fr or moore
output_lang	string	moore or fr
input	string	Source sentence
output	string	Target (reference) translation
reasoning	string	French explanation of the structural mapping (word order, compounds, absent copula, etc.)
word_by_word	list of {word, gloss}	Per-token French gloss of the Mooré side of the record
think	string	Raw model thinking span when present, else empty

Example (moore → fr):

{
  "input_lang": "moore",
  "output_lang": "fr",
  "input": "A Madaam Fatumata Wẽndengeta ... sẽn ya ẽspɛktɛɛr ...",
  "output": "Madame Fatoumata Windeguéta ... Inspecteur du trésor ...",
  "reasoning": "La phrase mooré suit l'ordre SVO, avec un nom composé suivi de titres et de fonctions ...",
  "word_by_word": [{"word": "madaam", "gloss": "madame"}, {"word": "ẽspɛktɛɛr", "gloss": "inspecteur"}]
}

How it was built

1. Source. 124 383 human FR↔Mooré pairs from burkimbia/fr_mos_annotated, each turned into two records (both translation directions).
2. Reasoning. Generated with Qwen3-32B-AWQ served by vLLM (offline batch inference on a single A100). The prompt grounds the model in a block of Mooré grammar rules (SVO order, serial verb constructions, unwritten tones, copulas yaa / bee / beeme, negation ka … ye, infinitive prefix n-) and asks for a structural French explanation only, so the model reasons about how the two sentences correspond rather than inventing lexical meanings.
3. Word map. word_by_word is not produced by the model. It is filled deterministically at aggregation time by glossing each Mooré token against a 20k-entry FR↔Mooré dictionary, with a small set of speaker-validated corrections taking priority. Unknown tokens are marked inconnu.

Usage

from datasets import load_dataset

ds = load_dataset("burkimbia/fr-moore-reasoning-v1")
train = ds["train"]            # 236 327 records
ex = train[0]
print(ex["input"], "->", ex["output"])
print(ex["reasoning"])
print(ex["word_by_word"])       # [{"word": ..., "gloss": ...}, ...]

# Keep only records with a full reasoning trace:
train = train.filter(lambda r: r["reasoning"])

Intended use

• Training / fine-tuning translation models that produce a reasoning trace.
• Studying structural correspondence between French and Mooré.
• A worked example of a reasoning-augmented corpus for a low-resource language.

Use in research

This dataset is released to support open research on low-resource machine translation and reasoning for African languages, in particular Mooré. It is suitable for work on reasoning-augmented translation, cross-lingual structural analysis, and evaluation of chain-of-thought on languages with little digital presence. If you use it in academic work, please cite it (below) and, where relevant, the source corpus burkimbia/fr_mos_annotated. We welcome feedback from Mooré speakers and researchers to improve gloss quality in future versions.

Limitations

• Reasoning is structural, not a proof. It describes word order, compounds, and grammar; it can still be imprecise or occasionally wrong on hard sentences.
• Glosses are dictionary-based. Mooré is tonal and tones are not written, so homographs exist (e.g. koom = eau vs kom = faim / enfant in compounds). Corrections cover the frequent cases; other glosses may be approximate, and unknown words are inconnu.
• ~1% of records have empty reasoning (very long inputs whose prompt was truncated). Filter on reasoning != "" if you need full coverage.
• Reasoning text is model-generated and inherits Qwen3's biases.

Citation

@misc{burkimbia_fr_moore_reasoning_v1,
  title  = {FR-Mooré Reasoning v1},
  author = {BurkimbIA Team},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/burkimbia/fr-moore-reasoning-v1}}
}