SFT

A fine-tuning phase in which a pre-trained model is trained on labelled demonstration examples to follow instructions. Supervised Fine-Tuning (SFT) is the first stage of alignment training for instruction-following LLMs, in which the model is fine-tuned on a curated dataset of (prompt, high-quality response) pairs to teach it to follow user instructions in the desired format.

SFT transforms a raw pre-trained model — which can predict text but has no concept of following instructions — into a model that consistently responds helpfully to user queries.

Supervised Fine-Tuning is a foundational concept in Training Technique. A fine-tuning phase in which a pre-trained model is trained on labelled demonstration examples to follow instructions.