Instruction tuning is a fine-tuning approach that trains language models on diverse instruction-response pairs, teaching the model to follow natural language instructions across many task types. The training data consists of examples formatted as: instruction describing what to do, optional input context, and the desired output. Tasks span question answering, summarization, translation, code generation, creative writing, analysis, and more. By training on thousands of different instruction types, the model learns the meta-skill of instruction following rather than just specific tasks. Instruction-tuned models generalize to new instructions not seen during training, understanding what humans mean by natural language commands. This is what transforms base language models from autocomplete systems into useful assistants. FLAN (Fine-tuned Language Net), InstructGPT, Alpaca, and Vicuna are landmark instruction-tuned models. The quality and diversity of instruction data directly impacts capability: more diverse instructions produce models that generalize better; higher quality responses produce models that give better answers. Self-instruct methods generate synthetic instruction data using the model itself, scaling instruction tuning beyond manually curated datasets. Instruction tuning often precedes RLHF: first teach the model to follow instructions, then refine behavior based on human preferences. Without instruction tuning, language models require careful prompting to elicit desired behavior.
Back to Glossary