koleshjr/multilingual_qa_training

build the multilingual health assistant Africa's communities deserve # Multilingual Health QA Training This repository contains the training and inference workflow used for a multilingual health question-answering competition. The core idea is to turn the labelled training data into a small retrieval knowledge base, enrich each example with similar labelled examples, fine-tune a chat model on those RAG-style prompts, and then run submission inference with vLLM. ## Final Performance This solution finished **11th place** on the Zindi leaderboard for the Multilingual Health Question Answering in Low-Resource African Languages Challenge. The current workflow is: ```text raw Train/Val/Test CSVs | v generate RAG context datasets with BGE-M3 retrieval | v fine-tune Sunbird/Sunflower with Unsloth + LoRA | v save/push checkpoint | v run vLLM inference with the LoRA adapter | v submission CSV ``` ## Why RAG-Enriched Fine-Tuning? We moved to RAG-enriched fine-tuning because the data showed strong evidence that retrieval is highly useful for this competition. The first clue was that a simple retrieval-only baseline did surprisingly well. Instead of generating answers with a large model, we retrieved the nearest question from the labelled data and copied its answer. Even simple TF-IDF retrieval scored well, and BGE-M3 embedding retrieval did much better on the leaderboard: ```text BGE-M3 retrieval: RougeL F1: 0.4823 Rouge1 F1: 0.5548 LLM Judge: 0.7379 ``` That means many test questions are semantically close to questions already present in train/validation. So the labelled dataset is not just training data; it is also a useful knowledge base. But retrieval alone is not enough. The best fine-tuned model still scored higher overall than retrieval-only, and first place was much higher. That suggests the task rewards both: ```text retrieval strength: finding close existing answers generation strength: adapting, rewriting, and answering novel or imperfect matches ``` So the idea behind RAG-enriched fine-tuning is to teach the model both b …