Instruction tuning fine-tunes a base model on a large collection of tasks expressed as instructions. Wei et al. (2022) showed in FLAN that this sharply improves zero-shot performance on unseen tasks, and Ouyang et al. (2022) combined instruction data with human feedback to build InstructGPT — the recipe that turned raw language models into helpful assistants.
It is the step that makes a model good at doing what it is told, rather than merely continuing text.