Fine-tuning Llama 3 on Clinical Notes: What Actually Worked

Lessons from fine-tuning a 70B model on 2M de-identified Indian clinical notes — data prep, QLoRA, and evaluation.