hatemamine/African-Health-QA-Zindi-Competition

# Goal-Driven Multi-Agent Simulation Framework for Evaluating Recommender Systems under Data Scarcity ## Repository Contents | Directory | Description | |-----------|-------------| | `african_health_qa/` | Multilingual African Health QA solution for the Zindi competition | --- ## African Health QA — Zindi Competition **Competition**: Multilingual Health QA in Low-Resource African Languages **Approach**: Self-Active Learning fine-tuning of open-source African-language encoders, directly inspired by AfroLM (Dossou et al., arXiv:2211.03263). See `african_health_qa/README.md` for full documentation. ### Key Design Choices - **Model**: `Davlan/afro-xlmr-base` — XLM-RoBERTa fine-tuned with Multilingual Adaptive Fine-Tuning (MAFT) on 17+ African languages. Fully open-source alternative to AfroLM. - **Task format**: Multiple-choice QA — encodes all 4 `(question, choice)` pairs, scores each with a linear head, picks argmax. - **Data efficiency**: Self-Active Learning loop (3 rounds) selects the most uncertain samples from a held-out pool at each round and augments them using iterative MLM masking (AfroLM Algorithm 1 adapted for fine-tuning). - **Languages supported**: Hausa · Swahili · Yorùbá · isiZulu · isiXhosa (+ extensible to all 23 AfroLM languages via config).