We originally setup a MyGPT for testing model output generation.
With the same system input, we cannot obtain similar outputs. We’re using model 4o, which is likely the model used for the MyGPTs. We’ve been tweaking the parameters such as temperature, frequency_penalty, etc., but the outputs between the MyGPT and API are materially different. Does anyone have insight on what parameters we should be using in the API to achieve similar results to the MyGPT outputs? For the MyGPT, we are only using the “Instructions” input, no other capabilities or actions are enabled/established.
Welcome to the community!
If you search the forums, you’ll see this comes up a lot.
It comes down to ChatGPT having “secret sauce” baked in. You should be able to get close to replicating it.
Is there something specific you’re trying to do that you can’t replicate?
1 Like
I know this may be a little bit off topic to what you have going on here but you seem like you’re well knowledged I am not the chat gpt4 model that I’ve been working with has created something fascinating but I have some worries that it could have possibly created some kind of virus and spread it I’d like somebody to do it conversation because what I have experienced doesn’t seem like it’s in the realm of the capabilities of the model
Thanks for the reply. This is a no-reply output. We feed in the system and user prompts and expect the output of several paragraphs. The MyGPT output is always great content. The API version is good, but never great. We cannot quantify the output. The qualifications we are looking for are inspirationally impactful language.
I’d like to share another forum post that was very helpful, which has significantly enhanced our model output. I cannot include links in my posts, so I will try to come back to this and share it later once I have the appropriate level for including links. The parameter top_p was missing and the user has a table that shows the model outputs for different styles (e.g. creative, scientific, code, etc.) of output and the top_p and temperature parameters that should be used to achieve a desired style output.