NasamuAlhassan/kusaal-mt

Fine-tuned NLLB-200 translation model for Kusaal (kus_Latn) ↔ English — MT system for this low-resource Gur language spoken in northern Ghana. # Kusaal ↔ English Machine Translation **An open-source machine translation model for Kusaal** — a Gur language spoken by ~400,000 people in northern Ghana and parts of Burkina Faso. Built by a native Kusaal speaker from Bawku, Ghana. Fine-tuned on a parallel corpus assembled from scratch, expanded through back-translation augmentation, and — as of August 2026 — a sentence-aligned corpus mined from Kusaal Wikipedia. **Model on Hugging Face** | **Live demo** | **Public benchmark** **In this repo:** the 13,659-pair Kusaal Wikipedia parallel corpus (CC BY-SA 4.0), the mining pipeline scripts, and the full pipeline report. **Licensing:** code and documentation in this repository are CC BY 4.0 (root LICENSE); the corpus/ directory is **CC BY-SA 4.0** (see corpus/LICENSE), inheriting the ShareAlike terms of its Wikipedia source text. --- ## August 2026 update: Wikipedia fine-tune The model was continued-trained on **13,659 new human-parallel sentence pairs** mined from Kusaal Wikipedia and its English counterparts (1,170 comparable article pairs, aligned by a shared-anchor + model-verification pipeline and partially human-adjudicated), plus a 6,917-sentence back-translation pool (real Kusaal, synthetic English, eng→kus direction only), mixed with the original corpus as replay. **Results** (beam 4, 1,000 sentences per test set per direction, sacreBLEU / chrF++ word_order=2): | Test set | Direction | BLEU before → after | chrF++ before → after | |---|---|---|---| | Wikipedia benchmark (new) | kus → eng | 42.17 → **47.73** | 59.92 → **64.09** | | Wikipedia benchmark (new) | eng → kus | 19.43 → **32.26** | 43.58 → **54.09** | | Original test set | kus → eng | 31.12 → 30.59 | 49.78 → 50.11 | | Original test set | eng → kus | 19.96 → 20.21 | 42.32 → 42.96 | The largest gain — **+12.8 BLEU for eng→kus on encyclopedic text** — comes from the back-translation pool targeting the model's weaker direction. Performance on the original (largely religious-register) test set i …