leny62/multilingual-health-qa

This project fine-tunes a multilingual T5 (mT5) model to answer health-related questions in five African languages: Akan, Amharic, Luganda, Swahili, and English. # Multilingual Health QA: Fine-Tuning mT5 for African Languages **Zindi Competition: Multilingual Health Question Answering in Low-Resource African Languages** **🎥 Watch the demo video** ## Overview This project fine-tunes a multilingual T5 (mT5) model to answer health-related questions in five African languages: Akan, Amharic, Luganda, Swahili, and English. The model is trained on a curated dataset of maternal, sexual, and reproductive health (MSRH) question-answer pairs across nine language-country configurations. Access to reliable health information remains a critical challenge across sub-Saharan Africa. Language barriers frequently prevent communities from receiving accurate health guidance in their native language. This project addresses that gap by building a multilingual model capable of generating fluent, accurate, and contextually appropriate health responses. ### Competition Metrics The model is evaluated using a weighted combination of: | Metric | Weight | Description | |--------|--------|-------------| | ROUGE-1 F1 | 0.37 | Lexical overlap of unigrams between prediction and reference | | ROUGE-L F1 | 0.37 | Longest Common Subsequence similarity | | LLM-as-a-Judge | 0.26 | AI-based evaluation of factual accuracy and completeness | An additional **AfroLM BertScore F1** metric evaluates top solutions using a multilingual transformer pretrained on 23 African languages. ## Project Structure ``` multilingual-health-qa/ |-- notebooks/ | |-- 01_eda.ipynb # Exploratory data analysis | |-- 02_finetuning.ipynb # Model fine-tuning pipeline (open in Colab) | |-- 03_inference_demo.ipynb # Evaluation, comparison, and demo |-- src/ | |-- config.py # Configuration and experiment definitions | |-- data.py # Data loading and preprocessing | |-- model.py # Model loading, LoRA setup, training | |-- evaluate.py # ROUGE metrics, experiment tracking | |-- ut …