Fine-tuning

In plain English

Fine-tuning takes a model that is already pre-trained and trains it further on a smaller, focused dataset so it performs better for a particular task, domain, or style. Unlike prompting, it actually changes the model weights, which makes the new behaviour durable. It is most useful for locking in a consistent style or format at scale, not for teaching the model new facts.

Why it matters

Fine-tuning can make a model reliably behave the way you need across thousands of requests, but it carries ongoing cost and maintenance, so it is worth it only when prompting genuinely is not enough.

A worked example

Common confusion

Fine-tuning shapes behaviour, not knowledge or reasoning. To give a model current or proprietary facts, retrieval-augmented generation is usually the better tool.