fine-tuning

In depth

The process

A fine-tuning starts from an already trained model and continues its training on a set of examples: inputs and the answers expected. On each pass, the weights are adjusted in small steps to bring what the model produces closer to those examples. The result is not a setting you switch on or off: it is a new model, to be stored, served, evaluated and versioned as such. More frugal methods train only a small part added to the original weights, which lowers the bill without changing the principle.

What places it on floor 1

That is what places fine-tuning on the floor of the model, whereas RAG belongs to the harness: one moves the weights, the other merely supplies the right documents to read at the moment of the question. The division of roles that follows is fairly stable. Fine-tuning excels at whatever is a way of doing: a tone, an output format, a trade vocabulary, a task of constant shape. It fails on whatever is a matter of substance, in particular facts that change, since no update happens without going through a training again. The cost, for its part, is paid once up front, and each request becomes a little cheaper since there is no longer any need to explain everything again in the prompt.

The costly trap

The costliest trap is wanting to teach facts by fine-tuning: you then teach the model to be wrong with more assurance, and the error becomes invisible because it can be tied to no source. A second trap is regression: in specialising a model, you often degrade general capabilities it possessed, which only shows if you also measure what you were not trying to improve. A third is maintenance debt, since the result is frozen on a starting version and every new generation of models raises the question of doing it all again. The prudent rule is to exhaust first what is reversible, the prompt and then RAG, and to fine-tune only once those levers have been used without the behaviour settling down.