Supervised fine-tuning
Also called Supervised Fine-Tuning
Training a model on labeled input-output examples to teach it a specific behavior or format.
Think of it like
Like showing an apprentice hundreds of worked problems with the correct answers until they reproduce the method.
Example
You collect 5,000 “customer message → ideal reply” pairs and fine-tune so the model matches your support tone exactly.
How it actually works
SFT is the straightforward flavor of fine-tuning: given inputs paired with the right outputs, minimize the error between prediction and target. It’s the backbone of instruction tuning and most task-specific adaptation. Its limits: it teaches imitation of the examples, so it can only be as good as the labeled data and can’t easily learn from “this is better than that” comparisons — which is where preference methods come in.
For product teams
The most direct way to shape behavior when you have clean examples of exactly what you want.
For engineers
Standard supervised learning on (input, target) pairs; the imitation step preceding preference optimization.
Related
- Fine-tuning — The general case is fine-tuning.
- Instruction tuning — Includes instruction tuning.
- RLHF — Complemented by RLHF.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome