Short Definition
Instruction tuning trains a model to follow natural-language instructions and produce useful responses.
Intuition
It teaches the model the interaction pattern users expect: read the request, follow constraints, and answer in the requested style.
Technical Definition
Instruction tuning fine-tunes a pretrained model on datasets of instructions paired with desired outputs, often before preference optimization.
Example
A base language model can be instruction-tuned to answer questions instead of merely continuing text.
Common Misunderstandings
Instruction tuning does not guarantee perfect instruction following.
It is different from prompting because it changes training, not only input text.