In plain English
Pretraining teaches a model language and facts by predicting the next token on huge text. Instruction tuning teaches it to follow requests. Feedback training shapes tone and safety. The model recalls patterns and has a knowledge cutoff, so give it facts when accuracy matters.
- Pretraining
- Learning by predicting the next token on huge text.
- Instruction tuning
- Training on examples of good answers.
- Knowledge cutoff
- The date after which the model learned nothing.
Picture it
How a chat model is made
1
Pretraining
Predict the next token on huge text
2