RichMePython/multilingual-health-qa-african-languages

AfriTeVa LoRA fine-tuning workflow for Swahili and Amharic health question answering # Multilingual Health Question Answering AfriTeVa Large LoRA fine-tuning workflow for Swahili and Amharic health question answering. The project targets the Zindi Multilingual Health Question Answering in Low-Resource African Languages challenge and optimizes generation settings using validation ROUGE. ## Contents - `afriteva_swahili_amharic_rouge_optimized.ipynb`: Google Colab notebook. - `fine_tune_afriteva_swahili_amharic_colab.py`: Colab-ready Python script. ## Data Download the challenge data from Zindi under its applicable terms. Place `Train.csv`, `Val.csv`, `Test.csv`, and `SampleSubmission.csv` together in the project directory, provide an extracted data directory, or upload the dataset ZIP when prompted in Colab. The raw challenge data is intentionally excluded from this public repository. ## Run In Google Colab 1. Upload the notebook or Python script to Google Colab. 2. Select a GPU runtime. 3. Configure `DATA_ZIP_PATH` or `DATA_DIR` if the data is already available. 4. Run the notebook cells or execute: ```python %run fine_tune_afriteva_swahili_amharic_colab.py ``` The workflow installs its pinned dependencies in Colab, mounts Google Drive, fine-tunes `castorini/afriteva_large` with PEFT LoRA, resumes from checkpoints, evaluates validation ROUGE, and writes test predictions.