I am trying to fine-tune a davinci model for experience sharing which is not working very well so far.
I found the section “More prompt advice” on OpenAI CookBook with wonderful advices. Tested on OpenAI playground page these advices have really improved the situation.
But to repeat all the advices into every prompts and completions in the jsonl training data lines is not realistic. So is there a better way?
When redo dataset still need a hit about how to do advices from More prompt advice like this one:
If you want it to say “I don’t know” when it doesn’t know the answer, tell it ‘Say “I don’t know” if you do not know the answer.’
Sounds like training data jsonl has a field other than ‘prompt’ and ‘comletion’ probably be ‘instruction’:''Say ‘I don’t know’ if you do not know the answer."?