A developer fine-tuned the Llama 3.2 3B language model to answer clinical questions in conversational, factual prose by cleaning and formatting a large real-world medical QA dataset (ChatDoctor HealthCareMagic 100K). The process involved removing platform-specific filler text, filtering low-quality or noisy samples, and converting data into a chat format suitable for Llama 3.2 training. This cleaning reduced the dataset from 112K to 45K high-quality samples, improving training signal quality. The cleaned dataset and fine-tuning pipeline are publicly available for reproducibility.

May 29, 2026, 7:13 PM