What is Fine-tuning?

What it is

Fine-tuning takes a model that has already learned general skills and trains it a bit more on examples from your use case. The model keeps what it knew and shifts its behavior toward the new examples.

Think of hiring an experienced writer and giving them your company's style guide and a stack of past articles. They are already skilled. You are teaching them your house habits.

How it works

A model starts with its parameters set by large-scale pretraining. In fine-tuning, you show it many examples of the input and the output you want, for instance support questions paired with ideal replies. The training process nudges the parameters so the model's outputs move closer to your examples.

Full fine-tuning updates all the parameters, which is expensive. Lighter methods, such as low-rank adapters, update only a small added set of numbers and leave the original model unchanged. They are cheaper and let you keep several variants of one base model.

Fine-tuning is one part of post-training, the stage that turns a raw text predictor into a useful assistant. Labs do it at scale. Developers do it on a smaller scale for their own products.

When it helps

• You need a consistent tone, format or structure that prompting does not hold reliably.
• You have a narrow, repeated task, such as classification, where a small tuned model can match a bigger general one at lower cost and latency.
• You want the model to follow a specialized vocabulary or workflow.

When it does not

Fine-tuning is poor at adding fresh facts. A model tuned on last month's documents may still get details wrong or forget them. For knowledge that changes, supplying the information in the prompt, within the context window, or retrieving it at question time is usually better. Fine-tuning also takes good data. A few hundred clean examples beat thousands of sloppy ones, and bad examples teach bad habits.

It can also cause a tuned model to lose some general ability, a problem called catastrophic forgetting, so test it on tasks outside your target.

Why it matters to you

Before you pay for fine-tuning, try a clearer prompt and a few examples inside it. If that gets you most of the way, you may not need it. If you do tune, measure the result against your own test cases, not just a public AI benchmark.

Related ideas

Fine-tuning is different from reinforcement learning, which learns from scored outcomes rather than from example answers. For open models, see open-weight model, since you can only tune weights you can download or that a provider offers to tune for you.