Loading…
Glossary term
Fine-tuning a model on labelled examples of the desired inputs and outputs.
In supervised fine-tuning (SFT) a model is trained on input and desired-output pairs using the usual next-token loss, but on a smaller, curated dataset.
It teaches format, style and task skills, and is usually followed by preference-based tuning.
SFT is the standard first step in adapting a base model for a purpose.