What is fine-tuning and when do you need it?

Large language models are first trained on a huge body of text. That stage is called pre-training and it is very expensive. Fine-tuning takes the general model and keeps training it on a smaller, specialised dataset. The model does not start from zero; it adds a new habit on top of what it already knows.

Think of a general practitioner who trains as a specialist. The basic knowledge stays in place and one kind of work gets better.

First you prepare example pairs: an input for the model and the correct output you expect. An insurer might use customer emails and the approved replies to them. The model is trained on these pairs and then tested on examples it did not see during training.

Today it is common to update only a small part of the model's parameters instead of all of them; the best known method is LoRA. These methods cut cost and hardware needs considerably.

In most cases, cheaper options come first.

• If the model misunderstands the task, improve the prompt and the examples you give it.
• If the model needs current or company specific knowledge, use RAG; no retraining is needed when the information changes.
• If you need a particular style, a fixed output format or a narrow task repeated many times, fine-tuning makes sense.

A fine-tuned small model can do that narrow job faster and cheaper than a large one.

Data quality decides most of the outcome: bad examples teach bad behavior. While focusing on the new task, a model can lose some general abilities, an effect known as catastrophic forgetting. When the base model moves to a new version, the fine-tuning usually has to be redone. Personal data in the training set can end up inside the model, so data selection needs to be planned with privacy law in mind from the start.