# What is Instruction Tuning?

Also called supervised fine-tuning, SFT.

Instruction tuning is the training stage that teaches a base model to follow written instructions and answer in a helpful assistant format. It uses a curated set of instruction and response pairs covering many task types. This stage converts a model that merely continues text into one that responds to requests, and it normally precedes preference-based alignment.

The training data pairs a request with a high quality response: summarize this passage, rewrite it formally, extract the dates, explain this error message. Coverage across many different task types matters more than sheer volume, because the goal is generalization to instructions the model has never seen rather than memorization of the particular ones it was shown.

Formatting conventions are learned at this stage too. Chat templates, the separation between a system message and a user message, and the habit of stopping cleanly once an answer is complete are all taught here rather than emerging naturally. Sending a prompt in the wrong template for a given model degrades quality noticeably, which catches out teams switching between families.

Instruction tuning is often confused with the feedback stage that follows it. Supervised instruction tuning shows the model what a good answer looks like by example. Preference-based training then ranks competing answers to refine tone, safety, and helpfulness. Most released assistants have been through both, and in casual use the two terms are frequently blurred together.

The stage is cheap relative to pre-training and is where a great deal of practical differentiation between products happens. It also introduces measurable side effects, including a tendency toward verbose, hedged, list-heavy answers, since responses of that shape were often labeled as high quality by annotators. Explicit prompting can partly counteract those learned habits.

## Key points

- Teaches a base model to follow requests instead of continuing text.
- Trained on curated instruction and response pairs across many tasks.
- Chat templates and system message conventions are learned here.
- Distinct from, and usually followed by, preference-based training.

## In practice

A base model given 'Write three subject lines for a product launch email' might reply with more example prompts, because that is a plausible text continuation. After instruction tuning on many thousands of request and answer pairs, the same input produces exactly three subject lines and then stops. Nothing about the model's underlying knowledge changed; only its response behavior did.

## Related terms

- [Fine-tuning](/en/glossary/fine-tuning)
- [Reinforcement Learning from Human Feedback](/en/glossary/reinforcement-learning-from-human-feedback)
- [System Prompt](/en/glossary/system-prompt)
- [Pre-training](/en/glossary/pre-training)
- [Prompt](/en/glossary/prompt)

[Back to the AI Glossary](/en/glossary)