Fine-tuning

Fine-tuning takes a model that's already trained and continues its training on your examples: hundreds or thousands of pairs of 'input we get' and 'output we want.' The model absorbs the pattern. It's how you get a model that classifies your support tickets exactly the way your team does, or writes in your house style without being reminded.

The common misconception is that fine-tuning teaches the model your facts. It doesn't do that well, and facts change anyway. For 'answer from our documents' the right tool is retrieval-augmented generation; fine-tuning is for behavior, not knowledge. The two combine fine: RAG supplies the facts, fine-tuning supplies the manner.

Practical example: a firm that has categorized 5,000 past inquiries has exactly the raw material fine-tuning wants. A small fine-tuned model can match their categories more cheaply and consistently than a large general model with a long prompt. That said, most small businesses never need fine-tuning, and we'll say so when that's the case. It earns its cost at volume.