Instruction Tuning (SFT)
Fine-tuning a pretrained language model on examples of instructions paired with good responses, turning a text-continuer into an assistant that follows requests.
Also supervised fine-tuning · SFT · instruction following · chat tuning · post-training
FLAN (Wei et al., 2021); InstructGPT (Ouyang et al., 2022)
A well-read person who has never had a job learns from a stack of example requests and model answers how to actually help customers.
Where it fits
+All relationships as a list
Is a kind of / part of
Siblings
Applies to
Used by
How it works
After pretraining, a model is good at continuing text but not at doing what it is asked. Instruction tuning, also called supervised fine-tuning (SFT), bridges that gap.
The data is pairs of instructions and ideal responses:
- “Summarize this email in two sentences.” → a good two-sentence summary
- “Write a Python function that reverses a string.” → clean code with a short explanation
- “What’s the capital of Australia?” → “Canberra.”
The model is fine-tuned on these, learning to predict the response given the instruction. Tens of thousands of high-quality examples can change behavior dramatically, because the knowledge is already there; SFT teaches the format and the role.
In the typical pipeline:
- Pretraining: learn language and knowledge.
- Instruction tuning: learn to follow requests.
- Preference tuning with RLHF, DPO or Constitutional AI: learn which of several acceptable answers people prefer, and what to decline.
Instruction tuning for a narrow task is often done cheaply with LoRA.
Everyday use cases
- Case 01
Turning a base model into a chatbot
Every chat assistant goes through this stage before preference tuning.
- Case 02
Company-specific assistants
Fine-tuning on internal question-answer pairs so a model answers in the company's format.
Strengths, weaknesses and when to use it
Strengths
- Simple supervised training with a big effect on usability.
Weaknesses
- Limited by the quality and diversity of the examples; can teach confident-sounding wrong answers.
When to use itWhen a pretrained model knows enough but does not respond in the way you need.
Compare with
Check yourself
- Q01
What does instruction tuning mainly change?
Show answer
A: How the model responds to requests