uenr-twi-malaria/nllb-twi-malaria

Twi Malaria Q&A — Fine-tuned NLLB-200

Fine-tuned version of facebook/nllb-200-distilled-600M for answering malaria-related health questions in Asante Twi, developed for a UENR final-year project. Used as the generation component in a retrieval-augmented (RAG) pipeline alongside ABENA for retrieval and ChromaDB as the vector store.

Training Data

• 1052 Twi malaria Q&A pairs, sourced from WHO, GHS, and CDC malaria fact sheets, translated to Twi and manually reviewed.
• Train/validation split: 946 / 106 rows (90/10).

Training Configuration

• Base model: facebook/nllb-200-distilled-600M
• Epochs: 10
• Optimizer: Adafactor
• Batch size: 1 (×8 gradient accumulation)
• Input format: Bua asɛm: {retrieved_context} Asɛm: {question} (matches the RAG pipeline's live inference format exactly)

Evaluation

Model	BLEU (held-out validation)
Base NLLB-200 (no fine-tuning)	12.22
Fine-tuned (this model)	99.16

Final training loss: 2.6490

Caveat: BLEU is an imperfect metric for a low-resource language with a small reference set. These scores should be read alongside qualitative human review, not as a standalone correctness guarantee.

Intended Use

General malaria health education in Twi, deployed behind a safety layer that enforces: malaria-scope-only responses, no diagnosis/prescription, emergency keyword referral, and a mandatory medical disclaimer on every response. Not a substitute for professional medical care.

Limitations

• Trained on a relatively small dataset for a generative model; answers on under-represented topics may be less reliable than well-covered ones.
• Twi translations in the training data were reviewed but Twi is a low-resource language for NLP tooling generally — treat outputs as a starting point for a health conversation, not a clinical source.

Deployment

Served via a Hugging Face Space using a RAG pipeline (ABENA retrieval + ChromaDB + this model for generation). See the Space for the live demo.

---

Model card auto-generated by the project's training notebook on 2026-09-01.