Modelpublic

Rekin226/nllb-3.3B-moore-lora-v0

sourceHugging Facecc-by-nc-4.0updated 2mo agoView on Hugging Face

0likes5downloads

Model Card

NLLB-200 Mooré LoRA (3.3B)

LoRA adapter for facebook/nllb-200-3.3B fine-tuned on ~205k cleaned Mooré↔French/English sentence pairs, all four directions (eng_Latn↔mos_Latn, fra_Latn↔mos_Latn).

Part of Mooré-Voice — open translation and speech recognition for Mooré (Mòoré / Mossi, ISO 639-3 mos), spoken by ~8 million people in and around Burkina Faso.

Evaluation

Model	Direction	BLEU	chrF++
zero-shot base	`eng_Latn→mos_Latn`	3.72	23.77
zero-shot base	`fra_Latn→mos_Latn`	3.18	22.82
zero-shot base	`mos_Latn→eng_Latn`	10.77	32.01
zero-shot base	`mos_Latn→fra_Latn`	9.14	29.86
fine-tuned	`eng_Latn→mos_Latn`	3.88	23.41
fine-tuned	`fra_Latn→mos_Latn`	3.5	23.32
fine-tuned	`mos_Latn→eng_Latn`	11.86	33.06
fine-tuned	`mos_Latn→fra_Latn`	11.24	32.66

Training data

Curated corpus v0.1 (see repo data/CORPORA.md): MT560 (Bible-register, ~89%), community instruction pairs, NLLB-mined bitext (LASER ≥ 1.15), translatewiki. Detokenised, LID-gated, FLORES-decontaminated. FLORES-200 devtest held out for eval.

Limitations

—Register skew: mostly religious text → weaker on administrative/technical register.
—Mooré orthography follows the 1976/2003 standard as used by the source corpora; diacritic usage varies upstream.
—Not human-evaluated yet; BLEU/chrF++ on FLORES only.

License note

Released CC-BY-NC-4.0 because a large share of the training text derives from sources whose redistribution terms are research-use-only or undeclared (see the repo's data/CORPORA.md / data/AUDIO_CORPORA.md). A fully permissive release is planned once the corpus is rebuilt on cleared sources (Common Voice mos + translatewiki + NLLB-mined).