I am unable to get the model to use the tool unless it’s specifically asked. I’ve tried to force it through instructions and tool names. When using normal gpt-4-turbo it does use the tool. This is a customer facing app, so I need it to not be asked, but the current bad behaviour from the gpt-4-turbo model is also unwanted.
Anyone else experiencing this?
Yes, I’m experiencing the same issue. The ability to call tools is very unstable in gpt-4-0125-preview. It doesn’t use tools very often, sometimes choose the wrong tool.
That doesn’t seem to make sense.
Documentation:
gpt-4-turbo-preview Currently points to
gpt-4-0125-preview.
The first thing I would do is reduce sampling parameters, such as top_p: 0.2, so you can keep the symptom straight and the output deterministic.
Then you may be able to find ultimately what the AI wants to do instead of random tokens being plucked that may send the course of output in multiple directions.
OpenAI of course doesn’t make it easy to diagnose improvements you could make: they turn off logprobs for functions.
That doesn’t seem to make sense.
Documentation:
gpt-4-turbo-preview Currently points togpt-4-0125-preview.
Well the original turbo - gpt-4-1106-preview. The model that had issues with tools as well.