Instruction tuning
Also known as: instruction tuning
Instruction tuning is fine-tuning specifically on examples of following instructions — “Summarize this…”, “Translate this…”, “Write a poem about…” with the ideal answer.
It’s what turns a model that only continues text into one that does tasks:
- Before: you feed it a question, it writes another question.
- After: you give a task, it completes it.
This is why assistants work: instruction tuning is the step that makes "helpful" the default mode. Without it, an LLM is just a very smart autocomplete.