For example, I use the following Prompt in the ChatGPT web version:
Goal:
Tell me a story.
ParaMeter:
temperature=2.0
presence_penalty=1
frequency_ penalty=0.5
In this Prompt, will the parameters used in the API such as temperature, presence_penalty, frequency_penalty be effective?
I think that after adding them, if I send them multiple times within a session, compared to a session without parameters, I do feel a change, but I am not sure if it is because the parameters are effective or because ChatGPT understands semantically that I need it to reduce repetitiveness.
The parameters are set by how the ChatGPT backend would call the model.
Not by the language within.
However, I have developed a classifier that will determine the type of output desired and set the API call’s temperature for recommended generation temperature. It would just be a bit more prompting to have it discover any user commands within that input wishing to change softmax logit temperature parameter directly.
Then your API-built chatbot could be instructed within language to do as you demonstrate.