监督微调（SFT）

Training a model on human-labeled instruction-response data to follow instructions.

SFT is the key step that makes a pretrained model "talk": high-quality dialogue data teaches answer format and task behavior, and it is the base alignment stage before RLHF.