Datasets:
yaaba-instruct
22,164 conversation turns in Mooré and French, the instruction data behind
yaaba-4B. Mooré (mos) is spoken by
about eight million people, mostly in Burkina Faso.
None of it is machine-translated. Every turn is built from written or transcribed Mooré: press, dictionaries, proverb collections, field interviews, a TV series and Bible text. Where a turn was assembled from a template, its content is still attested.
Tasks
| task | turns | what the model is asked |
|---|---|---|
traduire |
6,476 | translate between French and Mooré |
dire_en_moore |
5,749 | say a word or phrase in Mooré |
expliquer_proverbe |
3,063 | explain a proverb |
corriger |
2,967 | correct badly typed Mooré |
repondre |
2,798 | answer a question |
titrer |
615 | give a news article a title |
expliquer_texte |
249 | explain a passage |
raconter |
168 | tell a tale |
condenser |
79 | condense a text |
Fields
messages: one user turn and one assistant turn, in chat format.task: one of the nine tasks above.variant: the language of the question then of the answer, when the task has several (mos_mos,fr_mos,mos_fr,fr_fr), or the kind of turn.register:lexique,ecrit,entretien,enquete,presse,oral, and a few smaller ones.formulation: where the text comes from.attesteemeans both sides are taken from a document as is;assembleemeans an attested template filled with attested content;entree_degradeemeans the Mooré input carries a real typing error and the answer is the correct text;consigne_degradeemeans the instruction is typed carelessly (no accents, no quotes);validee_locuteurmeans signed off by a speaker;decideemarks the model's identity turns, written by the association.source: the document or collection the turn comes from.
Splits are drawn by answer, not by turn, so that the same proverb or the same title never appears on both sides.
Known limits
- Eight-digit numbers (Burkina phone numbers) are replaced by
XXXXXXXX; one TV scene reads a number out. - The lexical register dominates (54% of turns), because dictionaries are the largest clean source.
- A large share of the parallel material is religious.
- The division of dialogue into turns was inferred from layout and line order, and no speaker has reviewed that inference.
- Version of 24 September 2026. Compared with the set used to train yaaba-4B, it drops 996 correction turns whose target merely repeated the input, which taught the model to copy Mooré back instead of answering, and one title written twice in a row.
Licence and provenance
CC BY-NC 4.0: non-commercial use only, with attribution to BurkimbIA. BurkimbIA holds the rights to its text sources. Press text from raamde-bf.net is used and republished with the permission of its publisher; the TV series is used with the agreement of its production. Access is granted on request.
Built by BurkimbIA, a non-profit association in Ouagadougou building AI tools for the languages of Burkina Faso.
- Downloads last month
- 11