What is fine-tuning?

In this guide, you will learn what fine-tuning is, using the picture of an apprenticeship, and how it differs from handing the AI documents to read when you ask a question.

Starting from a capable generalist

Think of someone who has had a broad, excellent education. They read widely, write well, and can hold a sensible conversation on almost any topic. They are a capable generalist, but they have not yet specialised in your particular trade.

A general-purpose AI model is much like that. Trained on a huge sweep of text, it becomes a broadly capable starting point that many different tools can be built on top of. Because so much is built on it, this kind of broad, general model is called a foundation model.

The idea: a focused apprenticeship

Now imagine that capable graduate joins a law firm. Over months, they read the firm’s past cases, absorb its house style, and learn the particular way it handles clients. They were already smart. Now they are smart and specialised.

Fine-tuning does the same thing to a model. You take a foundation model and give it a further, focused round of training on a smaller, carefully chosen set of examples: your field, your tone, your kind of task. Teaching a general model to specialise like this is called fine-tuning.

Crucially, you are not building a new model from scratch. You are taking one that already understands language and steering it towards your speciality, which is far cheaper and quicker than the original training.

What it is good at

Fine-tuning shines when you want the model to consistently behave a certain way. A company might fine-tune a model on its past support replies so answers always land in the right tone. A medical group might fine-tune one on clinical notes so it handles their terminology naturally.

The examples you fine-tune on are a small, focused kind of training data. Quality matters more than quantity here: a few thousand clear, consistent examples of the behaviour you want will teach the model more than a mountain of messy ones.

The result is a model that has absorbed a style or skill so deeply it no longer needs to be reminded of it each time. The speciality becomes part of how it responds.

Fine-tuning versus giving it documents to read

Here is a distinction people often blur. There are two ways to make a model better at your world, and they solve different problems.

Fine-tuning changes the model itself, teaching it a lasting skill or style. Giving the model documents to read at the moment you ask, so it can answer from specific facts, is a different approach entirely, the one covered in what is RAG.

A rough rule of thumb: fine-tune to change how the model behaves, and reach for document-reading to change what it knows right now. If your facts change daily, reading fresh documents fits better, because fine-tuning bakes things in and cannot keep up with this morning’s figures. Many serious tools use both together.

Try it yourself

You cannot fine-tune a model from a chat window, but you can feel the difference a set of examples makes. Show a chat tool a few samples of a style, then ask it to continue in that style:

Here are three of my product descriptions. Study the tone, rhythm, and level
of detail, then write a fourth for a new product in the exact same style.

[paste three examples here]

New product: [describe it]

Notice how much closer the result sits to your voice once it has examples to learn from. Fine-tuning is this same idea, made permanent, so the model carries that style into every future response without being shown the examples again.

A useful clarification

A common worry is that fine-tuning means teaching the model everything from zero, which sounds daunting and expensive. It is the opposite. You are building on all the knowledge already baked into the foundation model and adding only a thin, focused layer of speciality on top. That is exactly why it is practical for smaller teams, not only the labs that trained the original.

Next steps

You have now seen how a general model can specialise. Next we look at AI that does more than talk: models that can take actions and use tools to carry out multi-step tasks, sometimes called agents.