Controlling the output

Question

Controlling the output

raihan0824 opened this issue 2 years ago · 3 comments

When the bot doesn't know/is not sure about the question, it replies with such things as: "I'm sorry, I don't have the knowledge that you ask". How do you guys achieve this? What kind of training data makes the bot respond like that? I'm highly curious about this.

Answer 1 · 2023-04-05T06:54:12.000Z

It learns from ChatGPT. All training data are available for download so feel free to examine yourself. Feel free to share your discovery here!

Answer 2 · 2023-04-05T08:48:40.000Z

so no pre/post process to do that? literally only relying on the training data?

Answer 3 · 2023-04-05T14:40:59.000Z

Well, there's a prompt for inference if that's what you mean by post-processing. See Appendix B in our paper.