Muen1/multilingual-health-qa-africa

# Multilingual Health Question Answering in Low-Resource African Languages ## Overview This is a fine-tuning pipeline for multilingual health QA across low-resource African languages, built for the Zindi Multilingual Health QA Challenge. The project systematically fine-tunes sequence-to-sequence models (mt5-small, flan-t5-base, flan-t5-large) using LoRA/PEFT, progressing through 6 experiments evaluated by ROUGE-1, ROUGE-L, and LLM-as-a-Judge. ## Repository Structure ```text multilingual-health-qa-africa/ ├── notebooks/ │ ├── 01_EDA.ipynb # Exploratory data analysis │ ├── 02_preprocessing.ipynb # Data cleaning and prompt construction │ ├── 03_finetuning_exp1and_exp2.ipynb # Exp 1 (zero-shot) & Exp 2 (mt5-small LoRA) │ ├── 03_finetuning_exp3.ipynb # Exp 3 (flan-t5-base LoRA r=16) │ ├── 03_finetuning_exp4_with_score_fixed.ipynb # Exp 4 (flan-t5-base LoRA r=32) │ ├── 03_finetuning_exp5_fixed.ipynb # Exp 5 (flan-t5-base prompt v2) │ ├── 03_finetuning_exp6.ipynb # Exp 6 (flan-t5-large LoRA r=16) │ └── 04_results_summary_fixed.ipynb # Aggregate results, plots, per-language ROUGE ├── data/ # Data files (stored in Google Drive) ├── submissions/ # Submission screenshots ├── plots/ # Charts and learning curves ├── requirements.txt └── README.md ``` **Note**: All data files and model checkpoints are stored in Google Drive at MyDrive/multilingual_health_qa/ due to Colab ephemeral storage. The notebooks mount Drive automatically in Cell 0. ## Setup Instructions (Google Colab) ### 1. Setup Google Drive Create the following folder structure in your Google Drive before running any notebook: ```text MyDrive/ └── multilingual_health_qa/ └── data/ ├── Train.csv ├── Val.csv ├── Test.csv └── SampleSubmission.csv ``` Download the competition data from Zindi and upload to that data/ folder. ### 2. Run notebooks in order Use T4 …