Instruction Tuning
Base models complete text. Instruction-tuned models follow tasks.
Instruction tuning uses examples like:
text
instruction -> ideal response
to teach the model how users expect an assistant to behave.
What instruction data teaches
- follow task wording
- answer directly
- use requested formats
- refuse unsafe requests
- ask clarifying questions
- use tools when examples include tools
Dataset quality
Good examples are:
- clear
- diverse
- correct
- formatted consistently
- reviewed
- representative of real use
Bad instruction data creates generic, verbose, or unsafe assistants.
Instruction tuning vs prompting
| Prompting | Instruction tuning |
|---|---|
| changes behavior per request | changes model behavior |
| fast to iterate | needs training data |
| no deployment burden | needs evals and rollout |
| good first step | good for repeated behavior |
Knowledge check
Q1: What does instruction tuning teach?
How to respond to tasks like an assistant.
Q2: Why not instruction-tune first?
Prompting and RAG may solve the problem cheaper and faster.