Supervised fine-tuning (SFT) is a post-training technique in which a pretrained base model is trained on conversation data consisting of instruction–exemplary response pairs. The data typically comprises tens of thousands to hundreds of thousands of examples written by human annotators or synthesized by other models, with each response ending in a stop token. A model trained this way is called an instruct or chat model; it acquires a consistent assistant persona and stops generating once its answer is complete.
Because SFT is supervised learning that imitates reference answers, it is distinguished from reinforcement learning methods such as RLHF and GRPO, which learn from reward signals. Typical LLM post-training first instills format and behavior through SFT and then refines preferences and reasoning ability with RL. In a narrower sense the term also covers fine-tuning in general, where a model is further trained on task- or domain-specific data.