Instruction Tuning: Definition and Examples

Instruction tuning (or instruction-based adjustment) is a key step in training large language models (LLMs). After pre-training on vast text corpora, the model has extensive linguistic knowledge but does not necessarily know how to respond usefully to a precise request. Instruction tuning bridges this gap by exposing the model to thousands of structured examples in the form of 'instruction → expected response'.

Concretely, a dataset is compiled consisting of various tasks: summarizing a text, translating a sentence, answering a question, generating code, rephrasing a paragraph, etc. Each example contains a clear instruction and the corresponding ideal response. The model thus learns to recognize the format of a command and to produce an output aligned with the user's intention.

One major benefit of instruction tuning is generalization: a model trained on a diverse set of instructed tasks becomes able to follow instructions it has never seen during training. This phenomenon makes models like ChatGPT, Claude, or Gemini so versatile from their launch.

Instruction tuning differs from RLHF (Reinforcement Learning from Human Feedback), which often comes after and focuses on aligning with human preferences. Both techniques are complementary: instruction tuning teaches the model to follow instructions, while RLHF refines the quality and safety of the responses produced.