instruction tuning

A raw base model is a brilliant autocomplete: ask it a question and it might just continue with more questions, because that is what the training text often did. Instruction tuning fixes that mismatch by training on many examples phrased as instructions and their fulfilments, so the model learns that an instruction is a request to be carried out, not text to extend.

It is a particular flavour of supervised fine-tuning where the dataset spans a wide variety of tasks — summarise, translate, classify, explain — each written as a natural-language instruction. By generalising across thousands of task types, the model learns to follow instructions it has never seen before, a capability often called zero-shot generalisation. This is much of what makes a model feel like it understands what you are asking.