Is there a way to play with Temperature, Top-P and Top-K on CustomGPTs?

Hi all,

I’m building a customGPT to teach some business related topic with a teacher’s IP. When I trying to test something like “Help my dog with his mortgage application” it actually gives me an answer, despite that on the instruction I ask to explicitly stay within the “realm of the real/possible”.

The reason I’m doing this test, is to make sure that the topics don’t get mixed up outside the scope and intention of the customGPT.

Someone recommended me to play with the Temperature, Top-P and Top-K, however, this only seem to be possible when using the Assistant API. Is there a way I can do this within the customGPT configuration? Maybe by putting it on the instructions?

Thank you all!

1 Like

API sampling parameters are not exposed to users in ChatGPT.

You might instead, for the concern you are facing, restrict the domain and purpose of the chatbot with clear instruction. Tell it what uses it shall accept or deny.

If a dog needed help with a mortgage, the first impediment is that the AI hasn’t been pretrained on dog language… :grin:

Thanks for the swift reply! :dog:

Just to clarify, this factors won’t work even if I specified them within the instructions?

Nope. Putting words in and getting words out cannot affect the algorithms that power that (unless you used a particular preclassification technique on your own API product, or even give the AI a function to change its own parameters on demand for future rounds).