Don't Fine-Tune. Unless You Can Answer These Three Questions.

Fine-tuning is the most expensive way to discover you had a prompt bug. I have gone looking for a retrain twice and found the problem somewhere else both times. Three questions to answer before you spend the money, and the one that stops most people: can you actually produce the data. My own build was silently discarding 83 of its best training examples, on every run, for months.