fine-tuning (an LLM)

A pretrained LLM has read a vast slice of the internet and knows language broadly, but it is a generalist. Fine-tuning is the short follow-up course that nudges that generalist toward what you actually want — a particular task, tone, output format, or body of knowledge. You keep almost everything the model already learned and adjust it gently with a much smaller, focused dataset.

Mechanically, you continue training the same network with gradient descent, but on your curated examples instead of raw web text. Each example produces a loss, gradients flow back, and the weights shift a little. Because the model starts from a strong base, it usually needs only thousands of examples and a small learning rate, not the trillions of tokens of pretraining. The result is a new checkpoint specialised for your use.

Fine-tuning is not the only way to adapt a model — prompting and retrieval often get you there with no training at all, and they are far cheaper to iterate on. Reach for fine-tuning when you need consistent behaviour, a fixed output format, or a style that prompting cannot reliably enforce.