How does a base model learn to answer us?

After pre-training, a model knows many patterns. But ask, "Give me three facts about Japan," and a base model may continue your question instead of answering it. Nobody has clearly shown it the job of being an assistant.

It is like hiring someone who has read a huge library but has never seen your workplace. They know a lot. They still need examples of what good work looks like here.

People prepare examples with two parts: a request and a good response. Request: "Give me three facts about Japan." Response: a short list of three accurate facts. The model studies many pairs like this.

This stage is often called supervised fine-tuning. The name is technical, but the idea is familiar. Show a new teammate examples of finished work, then let those examples teach the expected behaviour.

Give a new assistant a folder full of formal business letters and every message may begin with "Dear Sir or Madam." Give the same assistant short, friendly examples and the tone changes.

Training examples can teach formats, roles and specialist tasks. They also leave fingerprints on how the model sounds. More examples are not always the answer. Sometimes the model needs different examples.

Pre-training gave the model broad patterns. Demonstrations taught it how people expect those patterns to be used.

There can still be several reasonable answers to one request. The next step teaches the model which answers people tend to prefer.