Hi,

I'm currently conducting a small research on GPT-3 (using the API). The research involves some tasks which GPT should do; e.g. arithmetic with big numbers. The answers are always short and includes a single token. 

I would like to get multiple responses to the same question (mainly in order to do average and statistics); while GPT, of course, yields only one probability distribution, and won't change it if asked twice. Yes, there are cases where I can ask the same question with different objects (e.g. different numbers) but it's not always the case, and sometimes I would like to get different responses (and a lot of them) to the very same question.

I thought I could probably add noise before the real prompt, and maybe add a seperator between the noise and the real task (so gpt won't be too confuses). However, there are multiple ways to do it, and it's not clear which one is the best. I tried to google for it didn't find too much. Is someone here knows about someone who did it? academic papers would be best.

New Answer
New Comment

1 Answers sorted by

JNS

20

The API reference includes all the information you need, unless I am missing something.

"Temperature" or "top_p" for non determinism and "n" for multiple completions for the same prompt.