Instruction tuning is supervised post-training on instruction-response examples that teaches a pre-trained model to follow directions: turning a raw next-token
Supervised post-training on instruction-response examples that teaches a pre-trained model to follow directions, turning a raw next-token predictor into a model that answers questions, completes tasks, and respects formats.
It is the first step of post-training in virtually every modern assistant, preceding preference optimization like RLHF. The quality and diversity of the instruction data substantially shape the resulting model's behavior.