davidbirenzi/African-Language-Health-QA-Challenge

# Multilingual Health QA — Zindi Challenge Retrieval-first pipeline for the Zindi Multilingual Health Question Answering competition. Answers health questions in eight African language subsets by finding the best matching question in a labeled knowledge bank. **Best results:** Val weighted **0.4954** (EXP 17) | Zindi public **0.579346** (EXP 16) Demo Video: youtu.be ## Quick start ### 1. Clone and install ```bash git clone github.com cd African-Language-Health-QA-Challenge pip install -r requirements.txt ``` ### 2. Add competition data Download from Zindi and place in `data/`: ``` data/ Train.csv Val.csv Test.csv SampleSubmission.csv ``` ### 3. Run notebooks in order Open `notebooks/` and run from `01_eda.ipynb` through `10_experiment_exp16_rrf.ipynb`. Each notebook is self-contained and logs results to `reports/experiment_log.csv`. ### 4. Generate a Zindi submission **Best Zindi public score (EXP 16):** ```bash python scripts/run_exp16.py --submit ``` Output: `submissions/exp16_rrf_submission.csv` **Best Val model (EXP 17):** ```bash python scripts/run_exp17.py --submit ``` Output: `submissions/exp17_routing_submission.csv` --- ## Google Colab (mT5 fine-tuning) Fine-tune mT5 for generative experiments (EXP 12, 15): colab.research.google.com **Steps:** 1. Open `notebooks/colab_train_mt5.ipynb` in Colab 2. Set runtime to **T4 GPU** 3. Upload `Train.csv` when prompted 4. Download `mt5-health.zip` and unzip to `models/mt5-health/` 5. Run `notebooks/09_experiment_exp15_mt5_hybrid.ipynb` locally --- ## Project structure ``` African-Language-Health-QA-Challenge/ ├── data/ # Competition CSVs (not in git — download from Zindi) ├── notebooks/ # EDA + experiments 01–17 ├── src/ # Shared retrieval and experiment utilities ├── scripts/ # CLI runners for best submiss …