What is fine-tuning?
Taking a general model and training it further on narrower data to specialise it — cheaper than training from scratch.
Fine-tuning takes a model that has already completed its main training and trains it further on a smaller, narrower set of data chosen for a particular job. Rather than starting from nothing, the process adjusts a model that already handles general language well, nudging its internals so it performs better on the specific style, format or subject matter it’s being aimed at. Because the bulk of the learning is already done, fine-tuning is far cheaper and faster than training a model from scratch, which makes it practical for teams without the resources to build a model from zero.
When it is used
Fine-tuning earns its cost when a general-purpose model doesn’t do well enough on a specific job straight out of the box. A support team might fine-tune a model on years of past customer replies so it answers in the company’s tone and format automatically, rather than needing that context typed in every time. A research group might do the same with technical papers in one field, so the model picks up terminology it would otherwise handle clumsily. In each case, the goal is a narrower, more reliable specialist built on a broader foundation.
Not the only option
Fine-tuning is one route among several, and often not the first one worth trying: a carefully written prompt, or simply giving the model the relevant documents to work from at the moment of use — an approach known as retrieval-augmented generation, or RAG — can achieve the same practical goal without retraining anything at all. Products such as ChatGPT, and others including Claude, Gemini and Copilot, support forms of both approaches. Researchers at institutions such as Stanford HAI note that choosing between them involves trade-offs in cost, control and maintenance rather than a single correct answer.