← Explainers
2 min read

What is fine-tuning?

Teaching an existing AI model your style or task with extra training, and when you shouldn't bother.

BuzzwordsAI basics

Companies often say they “fine-tuned a model” for their product. It sounds like the answer to everything, but it’s one tool among several, and often not the first one to reach for.

Term
Fine-tuning
Extra training on your own examples

Taking a model that’s already trained and training it a bit more on your own examples, so it behaves in a specific way.

Think of it like this

Hiring an experienced writer and training them on your company’s style guide. They already know how to write; you’re teaching them your house style and the tasks you care about.

How it works

From general to specialised
  1. 01Trained model
  2. 02Your examples
  3. 03Extra training
  4. 04Specialised model
What you need
  1. 1Examples — many pairs of 'this input should get this output'
  2. 2Training run — the model's parameters are adjusted to match your examples
  3. 3Testing — check it improved and didn't get worse at other things

Which tool should you use?

Try them in this order
  • 1. Better promptsFast and free. Try first.
  • 2. RAGWhen it needs knowledge it doesn't have
  • 3. Fine-tuningWhen it needs a consistent style or behaviour

When fine-tuning makes sense

  • You need a very consistent format or tone across thousands of answers.
  • You have a narrow, repeated task, like sorting tickets into fixed categories.
  • You want a smaller, cheaper model to perform like a bigger one on that one task.

A common misunderstanding

“Fine-tuning is how you teach it facts.” It’s usually a poor way to add knowledge. Facts change and are hard to update this way. For knowledge, RAG is normally the better choice. Fine-tuning shines at teaching how to respond.

A typical case

Most companies do not need it to start. Prices and policies change often, and updating a document that RAG reads is quicker than retraining a model. Fine-tuning fits a fixed job better, like a hospital sorting notes into the same categories every day, or making replies follow one house style.

RAG · Parameters · Prompt