The training stage that turns a raw text predictor into something that follows instructions. A freshly pre-trained model, given a question, is as likely to produce a list of similar questions as an answer, because that is what it saw on the internet. Instruction tuning teaches it that a request should be met with a response. Almost everything people find useful about assistants starts here.
Technical view
Supervised fine-tuning on curated instruction and response pairs across a wide spread of tasks, typically a few tens of thousands to a few million examples, where quality and diversity matter far more than volume. It normally precedes preference-based training such as reinforcement learning from human feedback, which then refines tone, helpfulness and refusal behaviour. It also introduces the alignment tax, a small measurable drop on some raw capability benchmarks in exchange for usability.
Related
Part of AI Terminology Index, section: Building and Using Models.