Instruction tuning

Instruction tuning is fine-tuning a pretrained model on thousands of instruction-response examples, so it learns that its job is to follow instructions.

A freshly pretrained model is a scholar who has read the whole library but has never held a job. Ask it "Write a leave application" and it may continue your text — adding "for my cousin's wedding in October" — because continuing text is all it has ever done. Knowledge is present; the format of service is absent. Instruction tuning is the job training: example after example of the form here is a request → here is what a good assistant replies.

The data looks like this, tens of thousands to millions of times over:

instruction: "Explain photosynthesis to a ten-year-old."
response   : "Plants make their own food using sunlight..."

Technically it is ordinary supervised fine-tuning — which is why you will also see the name SFT, supervised fine-tuning. What changed history is what it teaches: not new knowledge but a role. After tuning, "Write a leave application" produces a leave application. The chat template — system-prompt, user turn, assistant turn — is learned here too.

In the standard assistant pipeline, instruction tuning is the middle act: pretraining builds capability, instruction tuning shapes the assistant behaviour, then preference training (RLHF or DPO) polishes quality and taste. Models on Hugging Face marked "-Instruct" have had this treatment; unmarked "base" models have not, and behave accordingly.

Where to go next#

• Full lesson: Fine-tuning
• Related terms: fine-tuning, rlhf, pretraining, system-prompt