Definition
Plain language
Teaching a model to follow what users ask by training it on lots of example instructions paired with good responses.
As stated in the literature
A post-training stage that fine-tunes a base language model on (instruction, response) pairs so it follows natural-language commands rather than merely continuing text; typically precedes RLHF and is implicated in side effects like calibration loss, mode collapse, and commitment sharpening.
Also called: instruction-tuned, instruction tuned, instruction-tuning
Why it matters: It turns a raw text predictor into something that follows directions, which is what makes a chatbot usable for ordinary requests.
For example, training a model on many pairs like 'Summarize this paragraph' followed by a good summary teaches it to actually do what users ask instead of just continuing the text.
Heard on the show
“That seed-and-amplify pattern is older than this paper — it's been used for instruction-tuning text data for years.”Episode 017 — When the Agent Grades Its Own Homework: A Brutal New Benchmark for AI Workers