Your SFT set is 20,000 expert-written answers to niche medical questions, and many cite facts the base model can't recall when you probe it closed-book.
After SFT the model matches the answer style, but on held-out questions it states confident, wrong specifics more often than the few-shot base model.