Davinci still seems like the gold standard, compared to turbo

[joao.occhiucci]

I figured out you are using max_tokens 4000. Try reducing it to get the
system + previous_conversation + response < 4000

[Securigy]

I will try that too, however, I also tried setting max_token: null, that is, using default value (which I am not sure what it is) and the result was the same Bad Request.

[ruby_coder]

@Securigy

max_token: null, that is, using default value

If you set max_token:null this will overwrite the client max_token value the client sends to the API, but the OpenAI hard limit of 4096 tokens cannot be bypassed or overridden by any client method.

HTH

See also related:

OpenAI API: chat completion pruning methods

Hi All, Hope all is well. I’ve been building out my test lab for the new chat completion API and it works well. The only real challenge has been the method to prune the messages in light of the very restrictive OpenAI API chat 4096 token limitation. This hard limit, as we devs know, generates errors and greatly restricts how developers can use the chat completion method, so I am currently experimenting with various “pruning methods” and if you have any suggests of other methods, please rep…

[bink]

Came across this thread and figured I’d chime in, even though the thread is dated. I too am getting the “As a large language model…” style responses more often than I care and I definitely consider this a bug, especially for an API. In my case, my system content is, consistently, no more than two relatively short sentences—so, as least for me, this has nothing to do with large system content.

Any new findings from anyone else here?