How an LLM is trained, stage by stage

How does a model that only continues text become one that answers you?

How LLMs work · about 5 minutes

A chat model starts as a base model that only continues text. Training runs in stages. The first stage reads a huge pile of text. The later ones teach the model to answer you, and to answer well.

Sketch: Xiaohei reads its way through a mountain of pages; an arrow leads to a small school desk and a blackboard marked "how to answer"