2

I want to do some tests using GPT-3. Instead of setting temperature = 0, I would like to use the best_of function. However, this gives me non-reproducible results since they differ in each iteration of code execution. Does anyone have an idea how I could achieve reproducible/deterministic results and still use best_of? Is there a way to use a random seed in the API for GPT-3?

Thanks!

Frigoooo
  • 51
  • 4

1 Answers1

0

Setting the temperature to 0 is the best way I'm aware of to achieve something approaching deterministic results.

You could also use fine tuning to push responses in a certain direction.

A final option is to "roll your own" via curation and maybe some type of chaining. If you want a certain response to always be the same, save it to a database and look it up. That might seem like a naïve answer, but solutions like this can be very sophisticated, especially with chain of thought style implementations.

Brian MacKay
  • 31,133
  • 17
  • 86
  • 125