Definition
Instruction tuning fine-tunes a pretrained language model on examples that pair a task description with a desired response. A training mixture may include question answering, classification, summarization, extraction, transformation, and other tasks expressed through natural-language instructions. The aim is to make the instruction itself a usable task specification, including for task types or phrasings not seen during tuning.
The method learns statistical behavior from demonstrations. It does not turn instructions into access controls, guarantee obedience, or establish which instruction wins when several conflict. Dataset wording, task balance, answer quality, and evaluation coverage determine what kind of instruction following the model acquires.
Term adoption
Jason Wei and colleagues' 2021 FLAN paper defined instruction tuning as fine-tuning language models on a collection of tasks described through instructions and used the term throughout. It is a clear early adoption point for the current label, but the paper does not establish that its authors were the first people to train models from natural-language task descriptions.
OpenAI's 2022 InstructGPT work used supervised demonstrations as its first training stage, then added preference modeling and reinforcement learning. That whole pipeline is broader than instruction tuning alone.
Distinguish it from nearby terms
Instruction tuning changes weights. A system prompt supplies runtime context. DPO and reward-model-based RLHF train from preferences rather than only desired demonstrations. An instruction-tuned model may later receive a system prompt, but the two mechanisms operate at different times.
Check your understanding
An instruction-tuned model follows formatting requests in ordinary prompts but ignores a rare conflict between a system rule and retrieved text. Which part calls for more training data, and which part requires an enforceable runtime boundary?