Sharif2138/African-Language-Health-QA-Challenge

# Multilingual Health Question Answering in Low-Resource African Languages All notebooks have a link to open directly in google colab Final project for Machine Learning Techniques I — Zindi Competition submission. This project fine-tunes multilingual sequence-to-sequence models (mT5, flan-t5) with LoRA adapters to answer health-related questions in low-resource African languages including Luganda, Amharic, Akan, Swahili, and English variants across Uganda, Ethiopia, Ghana, and Kenya. --- ## Competition **Zindi:** Multilingual Health Question Answering in Low-Resource African Languages Challenge **Evaluation metrics:** ROUGE-1 F1, ROUGE-L F1, LLM-as-a-Judge --- ## Repository Structure ├── Exp1.ipynb # EDA + preprocessing pipeline + retrieval baseline ├── Exp2_3.ipynb # mT5-small baseline + decoder start token fix ├── Exp4.ipynb # Proper baseline with sufficient data (500/lang) ├── Exp5.ipynb # Long instruction prompt experiment ├── Exp6_7_8.ipynb # Short prompt + LoRA rank ablation + mT5-base ├── Exp9_10_11.ipynb # Full dataset scale-up + flan-t5 + final mT5-base └── README.md --- ## Experiment Overview | Exp | Notebook | Model | Key Change | |-----|----------|-------|------------| | 1 | Exp1.ipynb | Retrieval (mpnet) | EDA + cosine similarity baseline | | 2 | Exp2_3.ipynb | mT5-small + LoRA r=4 | First generative attempt | | 3 | Exp2_3.ipynb | mT5-small + LoRA r=4 | Decoder start token fix + no_repeat_ngram | | 4 | Exp4.ipynb | mT5-small + LoRA r=4 | 500 samples/lang, true baseline | | 5 | Exp5.ipynb | mT5-small + LoRA r=4 | Long instruction prompt | | 6 | Exp6_7_8.ipynb | mT5-small + LoRA r=4 | Short prefix prompt | | 7 | Exp6_7_8.ipynb | mT5-small + LoRA r=16 | Higher LoRA rank ablation | | 8 | Exp6_7_8.ipynb | mT5-base + LoRA r=4 | Larger model on small data | | 9 | Exp9_10_11.ipynb | mT5-small + LoRA r=4 | Full dataset (29,751 samples) | | 10 | Exp9_10_11.ipynb | flan-t5-base + LoRA r=8 | Different mode …