temperature is at 0 so I can’t go lower. I’m changing my prompt because I was giving it very specific and long instructions to have it format responses in html, but now I’ll just get the json and do it myself.
Just yesterday, i was trying to transcribe some audio with Whisper. My temp was 0.9 or something like that, and it did return a lot of repeated lines from some point in the audio onwards.
Then, I changed to temp = 0. While it did improve the transcription a little (not everything from a certain point was repeating), what really made the transcription quite good was setting temp = 0.2 or so.
I think I have a very similar issue and wrote a detailed report.
On the GPT-3.5-turbo model messages were re-generated in a loop several times. Here is my issue / bug report. Maybe you want to check your usage and see if you have the same issue: