Spaces:
Running
Tool calling is inconsistent
Does the model that the user is interacting with, really have access to the tools or is it some other model which is used to detect if the user's request requires tool calling? I've noticed a lot of errors in tool calling where the model thinks that it has called the tool by writing something like "here are the images i generated" or "I searched the web" especially when the initial topic of conversation was something else. I think unlike other chatbots like copilot, Huggingchat models are not instructed to write specific command lines to call tools. I think that the active model should have real access to all the tools, for example copilot writes something like graphic_art("prompt") to generate images which is good because the model knows exactly what it's doing.
Yeah this is a known issue when prompting tool calls on llama 3.1 specifically, I find command R+ to be a bit better at it personally but it's not perfect for sure