instruction tuning

Training a model on lots of instruction-and-answer examples so it follows directions.

fine-tune on many (instruction, response) pairs across diverse tasks

Instruction tuning is a form of Book 01's fine-tuning where the curated examples are specifically (instruction, good response) pairs across many varied tasks — summarize this, classify that, answer this question. It teaches a raw pretrained model, which only knows how to continue text, to instead treat input as a request and respond helpfully. This is what unlocks strong zero-shot behavior, because the model generalizes the 'follow the instruction' skill to tasks it never saw. It usually comes before preference tuning like RLHF.

Appears in

Talking to it — prompting LLMs in the Wild · pt 2
Likely is not good Post-training · pt 1
Show it good answers Post-training · pt 2
Think longer at answer-time Post-training · pt 7

Nearby in the brain

LoRA base model preference alignment context window few-shot fine-tuning in-context learning prompt