What is fine-tuning?
Fine-tuning is when an existing AI model is trained further on your own examples to change how it answers.
A language model is first trained on huge amounts of general text. In fine-tuning it is then given a smaller set of selected examples. The model learns a format, a tone or a particular type of task.
Fine-tuning fits when the behaviour should change. One example is a model that should write summaries in a fixed structure.
For facts that change, fine-tuning is rarely the best route. A routine that is updated would need new training every time. It is usually simpler to let the model read current documents when it answers.
How it works
In fine-tuning the model is given pairs of instructions and desired answers. It responds, is compared with the example and nudges its weights. Often only a small part of the model changes, which makes it cheaper than training from scratch. The result is a new version of the model that needs testing before use.
Why it matters
Fine-tuning can make a model better at a narrow task, such as following a fixed template, using a particular professional vocabulary or classifying cases your way. It can also make a smaller model good enough, one that is faster and easier to run in your own environment.
What to keep in mind
The result is only as good as the examples, and poor examples teach the model the wrong things. A fine-tuned model also needs looking after. When a new base model arrives, the work often has to be redone.
So start by letting the model read your documents and giving it clear instructions. Fine-tune when that is not enough.
What it means for you
If you want AI to know your routines, letting it read them is usually enough. Fine-tuning is needed less often than people think.
Common questions
Should we fine-tune a model on our documents?
Rarely, if the goal is answers from those documents. That is done more simply with RAG, which reads current documents and can show sources. Fine-tuning suits changing tone or format better.
What happens to the data when a model is fine-tuned?
It becomes part of the model and cannot easily be removed. That calls for particular care with personal data and sensitive content.
May the supplier fine-tune on our data?
Only if the contract allows it. Ask explicitly whether your questions and documents are used to train or fine-tune models.
Related terms
Read more
All terms Updated