Fine-tuning

Further training of an existing model on your own examples, to change its behaviour rather than its knowledge.

Fine-tuning continues training a pretrained model on a dataset of your input and output pairs. It is effective for teaching a consistent format, a house style or a narrow classification task, and it can let a small model match a large one on that one task at much lower cost.

It is a poor way to add facts. Knowledge baked in this way cannot be updated without retraining and cannot be cited. For facts, use rag.

Before fine-tuning, exhaust prompting. A few worked examples in the prompt, described under few-shot-prompting, solve a surprising share of the problems people reach for fine-tuning to fix, with no training run and no dataset to maintain.

Where this comes up

Prompts, configs and tutorials in the library that touch this term.