Hello, we’ve been experiencing some issues with the API responses from the gpt-4-0125-preview model.
When making requests, we provide the chat history of the conversation with the user until that point and it’s last query, and a prompt that contains the necessary information to generate an adequate response.
However, the chat API is sometimes responding with randomized content, that has nothing to do with the user query or the prompt that is provided in the request.
Example response:
"
ies andbling and to the
abits
ia
iously
ocabulary
ines
ally andaticallyelame
and
inesamentevementvement
elines
el
ab and
abement
inasementsasm
and
and
el.
.
ab.ab and
and
and and the
and and
and
and
and
ils
and and
and
and and and and
and
and and and and and
and andellectualure
el Sam
2.5 Golf Club
3.5 Golf Club
4.5 Golf Club
5.5 Golf Club
6.5 Golf Club
7.5 Golf Club
8.5 Golf Club
9.5 Golf Club
…
"
When later trying to reproduce the bug by replicating the request, we will get the correct response from the API.
The configuration of our API request body is as follows:
“model”: “gpt-4-0125-preview”,
“temperature”: 0,
“max_tokens”: 1000,
“top_p”: 1,
“frequency_penalty”: 0,
“presence_penalty”: 0
Is there something that can be done to prevent this behavior from continuing?
Thank you!