Concept
Instruction Tuning
Supervised fine-tuning on (instruction, response) pairs to make a base model follow directions.
Definition
Instruction tuning turns a raw language model — which only predicts next tokens — into one that responds to prompts like a helpful assistant. The dataset is curated pairs of natural-language tasks and ideal completions across diverse domains. It is the step that bridges base models and chat models.
Common use cases
- Chat assistants
- Domain assistants
- Tool agents