Gradient
Lesson 10.1Concept~45 min

From predictor to assistant

Pretraining, fine-tuning and learning from human preferences.

Module 10 · 0/11 done

A predictor is not an assistant

The model you trained in the last module continues text. That’s all a base model does, no matter how big. Ask it “What is the capital of France?” and it may answer — or it may continue with three more quiz questions, because that’s what quiz pages on the internet look like.

Turning a raw predictor into a helpful assistant happens in stages:

Fig 10.1From raw text predictor to assistant. Click each stage to see what data it uses and how the same question gets answered.
Check yourself
During supervised fine-tuning, what is the training objective?

Complete the challenge above to continue.