Instruction tuning
Also called instruction fine-tuning, instruct model, chat tuning
Training that teaches a raw language model to follow instructions and answer like an assistant.
Instruction tuning is a stage of training that turns a base language model, which just continues text, into a model that follows instructions. The base model is trained further on many examples of instructions paired with good responses: summarise this, translate that, answer this question, write this code. After this training, it treats your text as a request to fulfil rather than a document to continue.
The difference is easy to see. Give a base model "Write a poem about autumn" and it might continue with "Write a poem about winter. Write a poem about spring.", treating the line as part of a list. An instruction-tuned model writes the poem. This is why open-weights models are often released in two versions, "base" and "instruct" or "chat".
Instruction tuning is usually followed by preference training, such as RLHF, which refines the model toward responses people rate as more helpful and safe. Together, these stages are what make zero-shot prompting work: the model has already seen countless instruction formats.
Example
When downloading an open model, you see "Model-8B" and "Model-8B-Instruct". For a chatbot or for following prompts like the ones on this site, you want the Instruct version. The base version is mainly for further training or research.
Common mistake
Running a base model and concluding the model is bad at following instructions. Base models are not designed to take instructions. Check that you are using an instruction-tuned or chat variant before judging quality.