You are building a medical QA assistant using SFT on do...

For each doctor-written response, you should probe the base model for the underlying factual claims before including the example in your training set. You must filter out examples where the base model is inconsistent or fabricates information on similar queries. It is also important to include explicit abstention examples for edge cases. Finally, consider a safer alternative by simplifying cited claims to general attributions rather than training the model to pattern-match specific author-year pairs it cannot verify.