noun · AI process
Fine-tuning
also called tuning
Extra schooling on your own examples. The AI keeps what it knows and learns to answer in your style or format.
Further training of an existing model on a smaller, task-specific dataset to adjust its behavior, style or format.
Training continues on a smaller set of example inputs and desired outputs, adjusting the existing weights so the model behaves the way those examples show. It changes style, format and habits far better than it adds reliable new facts.
Real-life example
You want every reply from your assistant in a strict house style. Long instructions work most of the time. You collect a few hundred ideal replies and fine-tune a model on them, and now the style comes out reliably with a short prompt. When your product facts change, though, you update documents it can read, not the model.
Where the analogy breaks
-
Fine-tuning is usually a poor way to add facts. It is better at style, format and behavior, while fresh knowledge is easier to supply through retrieval.
-
It needs good examples. A few hundred clean, consistent ones beat thousands of messy ones, and mistakes in the data are learned too.
-
It can make the model worse elsewhere. Narrow tuning may weaken general abilities, so test on more than your target task.