Instruction Tuning

The process of fine-tuning a base language model on a dataset of instructions and expected responses so it learns to follow user commands rather than just predicting the next word.

How it works

Instruction tuning (a form of SFT) trains a base language model on a large and diverse collection of instruction-response pairs covering many task types: summarisation, translation, question answering, coding, and more. The diversity of tasks is key — it enables the model to generalise the skill of 'following instructions' to novel tasks not seen during training. FLAN (Google), InstructGPT (OpenAI), and open datasets like Alpaca and OpenHermes use this approach.

Why it matters

Without instruction tuning, base models answer questions by continuing text patterns rather than actually answering. A base model asked 'What is the capital of France?' might output more questions rather than 'Paris.' Instruction tuning fundamentally changes the model's interface from 'text continuer' to 'instruction follower,' which is the prerequisite for any useful deployment. It is the step that made LLMs commercially viable as assistants and the foundation for all subsequent alignment work.