Open Gurkenglas opened 1 year ago
The strong theoretical reasons for prompts like QA QA QA Q don't hold when the model sees:
QA QA QA Q
user: QA QA QA Q assistant:
To fix this, try making it see:
user: Q assistant: A user: Q assistant: A user: Q assistant: A user: Q assistant:
(And I think OpenAI intends the Follow each Q with an A. prefix to be said by system.)
Follow each Q with an A.
system
Thanks for pointing this out.
I am keeping this ticket open until I find time for 2 things
The strong theoretical reasons for prompts like
QA QA QA Q
don't hold when the model sees:To fix this, try making it see:
(And I think OpenAI intends the
Follow each Q with an A.
prefix to be said bysystem
.)