Glossary

Instruction Model

An instruction model is the result of post-training, and it is meant to follow instructions. The name describes this goal: the model is meant to follow instructions that stand before it as text. It comes from the same base model, which was trained further with other examples and a different rating. The parameters are adjusted step by step in the same way, but the error is determined differently, as described under post-training.

An example: In the 2022 paper, the base model GPT-3 continued a task with further tasks. The result of post-training, InstructGPT, was preferred by the contractors over the answers of GPT-3 on a test set of prompts from held-out API customers who were not part of the training data, although it was clearly smaller with 1.3 billion parameters. What was measured was the preference in this test, not performance in all tasks. Exactly this adjustment to instructions is the goal of post-training.

Not to be confused with every chat model: Not every model that answers in a chat is an instruction model in the same sense. The term describes the goal of post-training, not the exact methods of a particular provider. The word also says nothing about whether an answer is correct.

Where you’ll come across it: In model names with an addition such as “Instruct”, for example Qwen2.5-7B-Instruct, and in descriptions of post-training.

Introduced in From Text Predictor to Assistant: What Post-Training Makes of a Model.

Last changed on · commit 653da06