Instruction-Tuned Model — AI Glossary

Reference details

Related terms

• Base Model
• Supervised Fine-Tuning (SFT)
• RLHF (Reinforcement Learning from Human Feedback)
• Direct Preference Optimization (DPO)

How it works

1. Supervised fine-tuning on instruction-response pairs teaches the model that a prompt is a request to satisfy rather than text to continue.
2. A preference stage then selects among acceptable responses, which is where tone, format habits and refusal behaviour come from.
3. The output distribution narrows as a result: the model becomes more predictable and less willing to produce continuations outside the assistant frame.
4. Chat templates formalise the roles, so the model expects turns in the structure it was trained on and degrades when that structure is violated.
5. The behaviour is learned rather than enforced, which is why instruction hierarchy is a tendency and not a security boundary.

Example: What the tuning narrows, and what that costs

Given an unfinished sentence, a base model continues it — which is exactly right for autocomplete and infilling. The instruction-tuned sibling is likely to explain the sentence or ask what you want done with it, because continuation is no longer its default frame. Same pretraining, different default. That is why a completion product often performs better on the base checkpoint, and why benchmarking a base model as an assistant measures the missing tuning rather than the model.

Commonly confused with

An instruction-tuned model follows directions; a base model of the same weights does not. It is also not the same as a reasoning model, which is trained to deliberate at length before answering. Chat model is a near-synonym in practice, usually implying the multi-turn message format on top of instruction tuning. Check which variant a model name refers to before comparing results.

When to use it

Reach for it when:

• Anything that should follow directions out of the box — assistants, tools, agents.
• Chat-formatted interaction, which the model was trained to expect.
• Explaining where refusal and formatting behaviour originate.

Reach for something else when:

• Raw completion or infilling, where the base model's continuation behaviour is the point.
• A specialised fine-tune, where prior tuning must be partly undone.
• Treating the instruction hierarchy as enforcement rather than a learned tendency.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Instruction-Tuned Model — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path