Instruction tuning
Fine-tuning a base model on example commands and good responses so it learns to follow directions.
Think of it like
Like taking a brilliant but literal new hire and running them through orientation until they reliably do what you ask, how you ask.
Example
A base model is trained on thousands of “instruction → ideal answer” pairs, turning raw text prediction into a usable assistant.
How it actually works
Pretraining leaves a model good at continuing text but not at obeying requests; instruction tuning fixes that with supervised examples of instructions paired with desired outputs. It’s usually the first alignment step, often followed by preference tuning like RLHF or DPO. It’s what converts a base model into something that answers questions instead of just autocompleting them.
For product teams
The step that makes a model act like an assistant; without it, even a huge model feels unusable.
For engineers
Supervised fine-tuning on (instruction, response) pairs to induce instruction-following before preference optimization.
Related
- Supervised fine-tuning — A form of supervised fine-tuning.
- Pretraining — Applied after pretraining.
- RLHF — Usually before RLHF.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome