What is fine-tuning?
A model's behaviour can be changed in three ways. Telling it, in the prompt. Showing it, by fetching the right text at the moment of asking. And training it: giving a provider thousands of examples of an input and the ideal output, and receiving a private copy of the model adjusted to match.
Training suits behaviour that is hard to describe and easy to show: a house style, a fixed format, a narrow job done millions of times. It can let a small, cheap model do one task as well as a large one.
It is a poor way to teach facts. Facts change, a trained model cannot say where it learned something, and it still makes things up. Retrieval is the tool for knowledge.
Training has real costs: collecting and cleaning examples, a fee, a test set to prove it helped, and doing it all again when the base model is replaced. Most teams reach for it last.
Why does fine-tuning matter?
It is the first thing many people ask for, and usually the last thing they need. Fine-tuning can make a model behave in a specific way without long instructions, and can make a small model do a narrow job as well as a large one. It is also slow, costly to maintain and the wrong tool for teaching facts. Knowing when it pays saves months.
How does fine-tuning work?
- Collect thousands of examples: an input and the ideal output.
- Check and clean them, because the model will copy their flaws.
- The provider trains a private copy of a model on them, adjusting its inner settings.
- Test the trained model against the original on cases it has not seen.
- Use it like any other model, and repeat the process when the base model is replaced.
When should you use fine-tuning?
Use it when
- A style or format is easier to show than to describe
- A narrow task runs at very high volume
- A small, cheap model must match a large one on one job
Think twice when
- The goal is knowledge that changes. Use retrieval
- A clearer prompt or a few examples has not been tried yet
- There is no test set to prove it helped
Where does fine-tuning sit in a real system?
Training happens at the provider, from examples you supply. The result is a model with its own name that your system calls. The examples and the test set are kept and maintained by you.
What goes wrong with fine-tuning?
- Training a model on facts and expecting it to stop making things up.
- Skipping the cheaper options: prompt, examples, retrieval, tools.
- Forgetting that every new base model means training again.
What does fine-tuning look like in practice?
A shop sorts a million messages a month into forty kinds. A large model does it for $4,000, right 94 times in 100. A small model trained on 50,000 messages, which the large one had labelled, does it for $200 and is right 95 times in 100.