Instruction tuning (SFT)
Beginner
Extra practice on examples of questions with good answers, so the model learns to be a helper.
Novice
Supervised fine-tuning on examples of instructions paired with good responses, written in a fixed template. It turns a base model that continues text into one that answers.
Expert
Next-token loss on prompt–response pairs, usually masked to the response tokens; tens of thousands of examples, small next to pretraining. A chat template marks the roles.
See also: Fine-tuning, RLHF.