Instruction Tuning
Teaching a model to answer in the shape a request calls for
- Instruction tuning teaches a model that already knows how to talk to answer in whatever shape a request calls for.
- Its training material is pairs of an instruction and the answer that fits it, usually written or cleaned up by people.
- A model that skipped this stage, asked a question, tends to string out more sentences like the question instead of answering it.
- This isn't where new knowledge goes in. It's where the model learns what shape to pull its existing knowledge into.
- After this stage comes another one, where human preferences polish the tone of the answers.
Contents
1The analogy
The night before a show goes up, the cast runs a rehearsal. Nobody teaches acting for the first time at this point — everyone already knows their lines and their emotional beats. What gets rehearsed is timing: entering when the director raises a hand, moving when the lights change, waiting when there's no cue at all.
An AI goes through something similar. Having read a mountain of text, it already knows how to string sentences together. But recognizing a person's request as a cue, and producing an answer that fits that cue, is a separate skill. So it goes through a stage where it rehearses timing using material that pairs an instruction with the answer that fits it. That's instruction tuning.
An actor doesn't come out of rehearsal a better actor. Same skill, just moving on cue now. A model that's been through this stage doesn't know more, either. What changes is when and in what shape it pulls out what it already knows.
2In detail
What the model looks like before rehearsal
A model that finished reading a mountain of text is a master of continuing text. Give it any sentence and it'll append a plausible next one. The trouble comes when what a person wants isn't a continuation.
Ask a model like this "tell me how to make chili" and, instead of a recipe, it might string out more requests in the same shape — "tell me how to make cornbread too," "and what side dish would go well with it." It didn't recognize the request as a request; it just continued the kind of text that tends to follow that sentence.
It's like an actor who delivers lines beautifully offstage but doesn't know the cue system yet. Not a lack of skill — a lack of knowing the rules of the show.
The rehearsal script is instruction-and-answer pairs
The material used for this training comes in pairs — an instruction a person might plausibly give on one side, a good answer that fits it on the other. Seeing a huge number of these pairs teaches the model the pattern: an answer of this shape follows an instruction of that shape.
Instructions cover a lot of ground: summarize this, put it in a table, rephrase this sentence, explain why, ask a follow-up because the request doesn't add up. Pairs where the right answer is "I don't know" for a genuinely hard question get mixed in too.
Variety matters more than sheer count. Load it up with only similar instructions and you get a model that's only good at that one shape — like rehearsing one scene twenty times and never touching the rest of the show.
It isn't where new knowledge goes in
It's easy to mistake this stage for injecting knowledge, but what actually grows is shape and manner. If you want the model handling something it genuinely doesn't know — company policy, a recent announcement — pulling in reference material alongside the question fits better.
Loading up on instruction examples from one particular field does make it comfortable with that field's tone and format, sure. But the effect on memorized facts is small, and if the examples you fed it contain mistakes, that same tone of mistake comes along too.
It's the same reason a rehearsal doesn't meaningfully improve an actor's acting. Blocking and cues stick firmly; the voice itself doesn't change.
A different layer from writing a good prompt
Writing a good instruction genuinely makes the answer better. That's like a hand signal given from the wings in the moment — it has to be repeated every time, and it disappears the instant a new conversation starts.
Instruction tuning is what lets the model move on cue without that hand signal at all. The values inside the model shift a little, so a short instruction alone is enough to get an answer in a reasonable length and shape. Not needing to attach examples every time to get a job done is a direct result of this stage.
Neither one replaces the other. Even a well-tuned model still responds strongly to a clear, specific instruction.
It leads into the next stage
By the end of this stage, the model follows instructions well, but the texture of its answers is still rough. When two answers both technically follow the instruction and one just reads much better, the standard for telling them apart wasn't taught here.
So one more stage follows. A person compares multiple answers, picks the better one, and that preference gets folded back in to refine the texture of the answers. This stage is called RLHF (Reinforcement Learning from Human Feedback).
Think of it as: rehearsal sets the timing first, then the pace and intensity of the performance get adjusted afterward based on how the audience reacts. The earlier stage handles shape; the later one handles polish.
3More precisely
Instruction tuning is supervised fine-tuning on instruction-and-response pairs. It nudges the values learned during pretraining only a little, without disturbing them drastically. The amount of data involved is tiny compared to pretraining — a good set of just tens of thousands of pairs commonly changes the result dramatically. These days, pairs written by people and pairs generated by a larger model both get mixed into the training set.
The analogy breaks down in a place worth naming. An actor understands and remembers the rules learned in rehearsal, but a model doesn't "understand" so much as have its values adjusted until an answer of a certain shape simply feels natural after an instruction of a certain shape. That's why it tends to follow instructions similar to ones it's seen even if they weren't in the script, while an instruction that departs sharply from that pattern can throw it off completely. And a rehearsal happens once, but this stage gets repeated with refreshed material every time a new model comes out. Handled carelessly, it can cost something too — training too long on too narrow a variety of pairs can dull the broad ability that pretraining built up. That's the reason people keep repeating that variety matters more than volume at this stage.
4Try it yourself
5Common misconceptions
It's easy to think instruction tuning adds new knowledge, but actually it's closer to learning what shape to pull existing knowledge into.
It's easy to think this is the same thing as writing a good prompt, but actually a prompt is a cue given in the moment, while this stage is an adjustment baked into the model itself.
It's easy to believe that after this stage the model will do absolutely anything asked of it, but actually the training material also includes situations where the right move is to ask a follow-up question or decline.
7One-line summary
In shortInstruction tuning is like a rehearsal that gets actors who already know their lines moving on cue — it's the stage that teaches a model to recognize an instruction as a signal and answer in the shape that fits it.
Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02