What is fine-tuning?
Teaching an existing AI model your style or task with extra training, and when you shouldn't bother.
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.
Taking a model that’s already trained and training it a bit more on your own examples, so it behaves in a specific way.
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
- 01Trained model
- 02Your examples
- 03Extra training
- 04Specialised model
- 1Examples — many pairs of 'this input should get this output'
- 2Training run — the model's parameters are adjusted to match your examples
- 3Testing — check it improved and didn't get worse at other things
Which tool should you use?
- 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.
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.
Related buzzwords
RAG · Parameters · Prompt