Function calling consistency
Hey,
Just curious if anyone can share how consistent they can get their instances to make structured function responses?
I've fed my instance the example weather function from the model card. It works pretty well until I give it some instructions before asking what the weather is. When prompting it for a greeting, then asking for the weather it hallucinates.
E.g:
User: Please give me the current weather for London
Assistant: Gives structured weather request
VS
User: Hi
Assistant: It seems like you're interested in the metadata for theget_current_weatherfunction.\n
User: Yes, please get the current weather for London
Assistant: Gives hallucination
Thanks for the comment, can you confirm how you are hitting the endpoint or querying the model? Are you copy-pasting the input from the card? or using apply_chat_template?
EDIT: I see that add_generation_prompt was missing from the apply_chat_template call, I've added that to the card now.