sashishjha/Multilingual-Health-Question-Answering-in-Low-Resource-African-Languages-Challenge-by-ITU

# AfriQA — Multilingual Health QA Competition **Goal:** Score 0.72+ on Zindi leaderboard | **Deadline:** June 22, 2026 ## Quick Start (Server) ```bash cd /mnt/data/sashishj/projects/afriqa sbatch slurm/run_exp10_best.sbatch # THE main experiment tail -f logs/exp10_best_*.log # Monitor ``` ## What This Does **Experiment 10** (our best combo): 1. **Phase 1:** Trains `mT5-Large` (1.2B) with LoRA on merged train+val (36k samples), 3 GPUs, 5 epochs 2. **Phase 2:** Runs RAG inference (BM25 retrieval + generation) using the trained model 3. **Output:** Two submission CSVs in `submissions/` ## Repo Structure ``` afriqa/ ├── configs/ │ └── experiment_10_best.yaml # THE config (mT5-Large + merged + RAG) ├── scripts/ │ ├── run_exp1_generation.py # Main training script (all experiments use this) │ └── run_exp3_rag_baseline.py # RAG inference script ├── slurm/ │ ├── run_exp10_best.sbatch # THE sbatch to run │ └── run_exp6.sbatch # Previous mT5-Large experiment (kept for reference) ├── src/ │ ├── data_loader.py # CSV loading + tokenization │ ├── metrics.py # ROUGE scoring │ ├── utils.py # Seed, config, logging helpers │ ├── training/trainer.py # LoRA + Seq2SeqTrainer setup │ ├── inference/predictor.py # Batch prediction │ ├── evaluation/evaluator.py # Per-subset evaluation │ ├── retrieval/ # BM25, Dense (LaBSE), Hybrid retrievers │ └── rag/pipeline.py # RAG: retrieve examples → generate answer ├── notebooks/ # (empty — Kaggle/Colab were unstable) ├── submissions/ # Output CSVs for Zindi ├── Train.csv / Val.csv / Test.csv # Competition data ├── SampleSubmission.csv # Zindi format reference └── requirements.txt # Python dependencies ``` ## Scoring History | # | Experiment | Score | Key Insight | |---|---|---|---| | 1 | mT5- …