Using v2 Model Calls Functions with Hallucinated Values

[meetkai]

Using v2 Model Calls Functions with Hallucinated Values

[musabgultekin]

Thanks for pointing that out. Its probably due to the new dataset mix. Will check that out

[sashokbg]

Hello, @musabgultekin I would like to propose something. Please bear in mind that I am a complete beginner in LLM and ML.

What if we add a parameter to the functions definition json that defines what functions should never be called with invented values ?

Let's say if we put "requires_user_validation": true and train the model to never call this function without prompting the user for all the parameters.

[rgbkrk]

Would you all be open to some new synthetic datasets? I'm going to start gathering some GPT-4 usage with computational tools as datasets performing deeper analysis work. Any idea what formats would be best for you? The simplest way for me to store them for now is just the literal OpenAI messages format.