Fine-tuning means further training an already pre-trained language model on your own collection of examples. The aim is not to teach the model new knowledge but to shape its behaviour: a particular form of answer, a consistent style, a reliable output format, or confident handling of a specific task.
What it is actually good for
The value of fine-tuning lies in form, not content. Where a model keeps deviating from the desired format despite careful instructions, where a professional register must be hit consistently, or where a narrowly defined classification task needs high reliability, fine-tuning is the right instrument. A further, often overlooked effect: a fine-tuned model needs shorter instructions, because the behaviour sits in the model rather than the prompt — which can lower cost per request.
What it is not good for
Fine-tuning is the wrong tool for keeping knowledge current. Train facts into a model and they are outdated at the next change, and correcting them means training again. Facts that change belong in a queryable source, not in model weights. It is equally unsuited to confidential or tenant-specific data: what has been trained into a model cannot be selectively deleted again.
The relationship to RAG
The question is often posed as either-or and is a both-and. RAG supplies the knowledge, fine-tuning shapes the form. A system that must cite current documents correctly while answering consistently in a given format benefits from both. In sequence, though, it is almost always: RAG and careful prompting first, fine-tuning after — because the first two are cheaper, faster to change, and in most cases already sufficient.
The data question
The effort of fine-tuning lies not in the training but in the data. What is needed is a collection of examples showing the desired behaviour correctly and consistently. Inconsistent examples produce inconsistent behaviour — the model learns exactly the contradictions present in the data. Quality clearly beats quantity here: a small, carefully curated collection generally achieves more than a large, automatically assembled one.
The ongoing costs
A fine-tuned model is not a one-off artefact but a dependency. It must be maintained, evaluated, and rebuilt when the base model changes. Those running costs are regularly underestimated before the decision and usually exceed the training cost over the lifetime.
Practical consequence
The useful order is: careful prompting, then RAG, then — if the form still does not hold — fine-tuning. Starting the other way round invests early in a dependency that a better prompt might have made unnecessary.