AI Glossary Term
What does “Instruction Tuning” mean?
Fine-tuning a model on instruction-response pairs to improve task following.
How Instruction Tuning works
Example
A raw pre-trained model given 'Summarize this article' might just continue the text. After training on many examples of instructions paired with good responses, it learns to actually produce a summary.
Why it matters
Instruction tuning is a key step that turns a text predictor into something that responds to requests. It is usually followed by preference training to refine behavior.
Common questions
Where does the training data come from?
Human-written examples, existing datasets reformatted as instructions, and sometimes model-generated examples that are filtered.
Is it the same as RLHF?
No. Instruction tuning uses supervised examples; RLHF uses human preference ratings. They are often combined.
Does it add knowledge?
Mostly it changes how the model responds, not what it knows.