supervised fine-tuning (SFT)

Supervised fine-tuning is teaching by worked examples. You collect pairs of prompt and ideal response — questions paired with the answers you wish the model would give — and train the model to reproduce those answers. It is called supervised because every example carries a target the model is graded against.

The model reads the prompt plus the gold response and is trained with the same next-token objective as pretraining, except the loss is usually computed only on the response tokens, not the prompt. Over many such pairs it learns the mapping from instruction to good answer. SFT is the first stage of turning a raw base model into a helpful assistant, and it typically comes before preference-based alignment such as RLHF.