04Book 04 · Chapter 04 of 07· 5 min · Instruction following
This is a question, not a blog
Instruction examples are in chat form: a user request, then the reply the model should give. After this stage, a question is treated as something to answer, not as a blog post or forum thread to keep writing.
Raw models continue text. Paste a question and they may write another question, a comment thread, or a 400-word article that never provides an answer the user wants or can use.
Instruction examples are in a chat shape: a request, then a target reply. The model practices that shape until a question usually gets an answer.
The quality of those answers still depends on the quality of the training examples. Short true ones teach short-and-true. Long sure ones teach long-and-sure. “Canberra.” teaches “Canberra.” “Well, many people visit Sydney…” teaches a waffle.
User
What is the capital of Australia?
Assistant
What is the capital of New Zealand? What is the capital of Canada? Reply if you agree Sydney is nicer though.
continues text. A question can grow more questions.
Quiz
6 questions
Before moving forward, test your knowledge of what you’ve just learned!
1.Without instruction examples, a question may be treated as:
2.Instruction tuning uses:
3.The quality of instruction replies:
4.A user message after this stage is usually treated as:
5.Long sure examples teach:
6.This stage still uses:
0 of 6 answered
4 / 7