When Not to Fine-Tune

Fine-tuning is often the first idea and rarely the right one. It is worth knowing which problems it solves before committing to the cost of a training run and a dataset.

Knowledge problems are not fine-tuning problems

If the model does not know your prices, policies or product catalogue, retrieval is the answer. Fine-tuning teaches behaviour and format far better than it teaches facts, and facts embedded in weights are difficult to update and impossible to cite.

Try prompting properly first

A large share of fine-tuning projects are launched against a mediocre prompt. Clear instructions, a handful of well-chosen examples in the context, and a structured output schema close most of the gap, in an afternoon rather than a month, and remain easy to change.

When it genuinely wins

A consistent tone or format you cannot reliably obtain through prompting, a narrow classification task where a smaller tuned model is dramatically cheaper at volume, or a specialised domain vocabulary. In each case you need a few hundred to a few thousand good examples.

Establish a baseline with prompting and measure it. Without that number you cannot tell whether the fine-tune helped, and most teams never find out.