Fine-tuning

Fine-tuning continues the training of an existing model on your own examples, adjusting its weights so it behaves more like those examples. It is how you teach a model a house style, a specific output format, or a classification task with your own categories.

It is regularly proposed for the wrong problem. Fine-tuning teaches behaviour, not facts — a model fine-tuned on your product documentation will learn to sound like your documentation, while still inventing details. For facts that must be current and correct, retrieval-augmented generation is the answer.

It also carries an ongoing cost that a demo does not reveal: the tuned model is a version you now own and must re-tune when the base model is superseded, along with keeping the training set that produced it.