166
fine-tuning
Continue training on your examples. It does not replace a missing fact, and it is not the first thing to try.
What is fine-tuning?
Fine-tuning adjusts a model’s weights on a narrower dataset so style, format, or a task shifts. It memorizes what is in that set. Facts that change (prices, policies) belong in retrieval or tools, not in weights. Prompting and few-shot come first because they are reversible.
Why does fine-tuning matter when vibe coding?
Models offer to fine-tune when the prompt is underspecified. You pay to bake in today’s wording. Name the cheaper step first.
How do you do fine-tuning?
Fix the prompt and the examples. If format still fails on an eval set, then consider a tune — on a frozen eval you did not train on. Keep a way to roll back the weights.
How do you ask a model for fine-tuning?
Do not fine-tune yet. Fix the prompt and add few-shot examples. If we tune, hold out an eval set and do not train on it. Do not fine-tune to store facts that change; those stay in retrieval.
What goes wrong with fine-tuning?
Tuning on the same twenty demo prompts you evaluate on. The score is memory.