Fine-Tuning

Pretraining gives you a model that knows a lot but doesn't behave like an assistant. Fine-tuning adapts that base model to a specific job or style.

The method is the same training loop, run on a smaller, higher-quality dataset aimed at what you want. To make a support assistant, you'd fine-tune on real support transcripts. The model's parameters shift a little, enough to specialize without forgetting what pretraining taught it.

A common first step is instruction tuning: fine-tuning on examples of instructions paired with good responses. This is what turns a raw next-token predictor into something that follows directions and answers questions.

Fine-tuning everything is expensive, so practitioners often use lighter methods like LoRA, which trains a small set of extra weights and leaves the original model frozen. You get most of the benefit at a fraction of the cost.

Fine-tuning teaches from fixed examples, which is supervised learning: here's the input, here's the right answer, match it. As the first stage of post-training you'll see it written as supervised fine-tuning, or SFT, the step that comes before reinforcement learning takes over to shape the behavior that's hard to write down as right answers.