SamAbr/Multilingual-Health-Question-Answering-in-Low-Resource-African-Languages

# Multilingual Health Question Answering in Low-Resource African Languages This repository contains code, notebooks, and models for building end-to-end question-answering systems in low-resource African languages (Akan, Amharic, Luganda, Swahili, and regional English variants). --- ## 📁 Directory & Project Structure ``` Multilingual-Health-Question-Answering-in-Low-Resource-African-Languages/ ├── data/ │ ├── raw/ # Original raw Zindi datasets │ │ ├── Training set.csv # Training dataset │ │ ├── Validation set.csv# Validation dataset │ │ └── Test set.csv # Test evaluation dataset │ └── processed/ # Preprocessed & augmented datasets ├── notebooks/ # Jupyter Notebooks for analysis & cloud training │ ├── colab_nllb_health_qa.ipynb # Cloud/Colab automated clone & run notebook │ └── multilingual_health_qa_starter_notebook.ipynb # Interactive analysis & hybrid modeling notebook ├── src/ # Modular Python scripts & pipeline algorithms │ ├── __init__.py │ ├── nllb_pipeline.py # NLLB-200 generation, per-language batch sampler & callback │ ├── retrieval.py # TF-IDF, Multilingual Dense Sentence Embeddings & Hybrid RAG │ └── threshold_optimizer.py# Per-subset threshold grid optimization & score calibration ├── models/ │ └── checkpoints/ # Model checkpoints & fine-tuned weights ├── submissions/ # Formatted CSV submission outputs for Zindi ├── LICENSE └── README.md ``` --- ## 🚀 Getting Started ### 1. Requirements & Dependencies Ensure you have PyTorch, Transformers, Sentence-Transformers, Datasets, and evaluate installed: ```bash pip install torch transformers sentence-transformers datasets evaluate scikit-learn pandas numpy rouge-score peft accelerate ``` ### 2. Running in Cloud / Google Colab / Docker Environments Open `notebooks/colab_nllb_health_qa.ipynb` in Google Colab or your cloud Jupyter notebook server. It inclu …