Ask a raw base model "What is the capital of France?" and a plausible response is: "What is the capital of Germany? What is the capital of Spain?" It is not being unhelpful. It was trained to continue text, and a question in a document is frequently followed by more questions.
Everything that makes a model feel like an assistant — answering rather than continuing, following an instruction, stopping when finished — is added afterwards. Instruction tuning is the first and largest part of that.
What changed in 2026
- The stage boundaries blurred. Instruction and preference tuning increasingly happen in combined pipelines rather than as cleanly separated phases.
- Data quality decisively beat quantity. Carefully curated, smaller instruction sets consistently outperformed large scraped ones.
- Synthetic instruction data matured. Generating and filtering instruction pairs with a stronger model became standard, with contamination checks alongside.
- Base models stayed available. Access to non-instruction-tuned checkpoints kept custom instruction tuning viable rather than purely academic.
What it actually does
The training data is pairs: an instruction, and the response you want. The model learns, through ordinary supervised training, to produce that kind of response to that kind of input.
Three behaviours emerge that a base model lacks:
Responding rather than continuing. The model learns that an instruction is followed by an answer, not by more instructions.
Format adherence. Asked for three bullet points, it produces three bullet points, because the training data consistently did.
Stopping. Base models run until they hit a token limit. Instruction-tuned models learn that responses end, which is why they emit an end token at a sensible point.
None of this adds knowledge. The facts were learned during pretraining; instruction tuning changes how they are surfaced.
Diversity beats volume
The finding that has held up repeatedly: a smaller, more varied instruction set outperforms a larger, narrower one.
A dataset of fifty thousand examples that are all summarisation teaches summarisation and generalises poorly. A few thousand examples spanning summarisation, extraction, reasoning, rewriting, classification, refusal, and multi-turn conversation teaches the general shape of following instructions, which transfers.
| Dataset property |
Effect |
| Many task types |
Strong generalisation |
| High response quality |
Sets the ceiling on output quality |
| Consistent formatting |
Reliable format adherence |
| Includes refusals |
Model learns to decline appropriately |
| Includes multi-turn |
Conversation coherence |
| Sheer size alone |
Weak predictor of quality |
The response quality row deserves emphasis: the model learns to imitate what it sees. Mediocre responses in the training data produce a mediocre assistant, regardless of how many there are.
Where it sits in the pipeline
Instruction tuning teaches the model to produce a reasonable response. It does not teach it which of two reasonable responses is better — that is what preference tuning adds afterwards, using comparisons rather than demonstrations.
The ordering matters. Preference methods refine a model that already responds sensibly; applying them to a base model that still continues text has little to work with. See DPO vs RLHF for the second stage.
Task-specific fine-tuning is a third, separate thing. If you start from a model that is already instruction-tuned and train it on your particular task, you are not instruction tuning — you are specialising, usually with an adapter, per QLoRA.
Common mistakes
- Instruction tuning an already-tuned model. You want task specialisation, on a different dataset.
- Optimising for dataset size. Diversity and quality dominate.
- Low-quality responses in the data. The model imitates them faithfully.
- No refusal examples. Produces a model that attempts everything, including what it should decline.
- No multi-turn examples. Single-turn training produces a model that loses conversational thread.
- Skipping contamination checks on synthetic data. Generated examples can leak evaluation content.
FAQ
Can I instruction-tune with a small dataset?
Yes, and this is one of the areas where a few thousand carefully-built examples genuinely works. The constraint is curation effort, not volume.
How do I evaluate it?
Held-out instructions the model never saw, scored on whether the response actually follows the instruction rather than on similarity to a reference. See multi-turn evals for the conversational side.
Is this the same as prompt engineering?
No — prompting works within a model's existing behaviour, instruction tuning changes that behaviour by updating weights. Prompting is cheaper and reversible, which is why it is the right first attempt.
Do I need a base model?
For genuine instruction tuning, yes. Most teams should start from an instruction-tuned model and specialise, which is a different and much cheaper task.
Where to go next
For the preference stage that follows, read DPO vs RLHF and reward models. For the efficient way to specialise afterwards, QLoRA, and for building the data, data curation for LLMs.