Sponsored by: VAGO Solutions and HyperSpace.Ai
A function calling version of cognitivecomputations/dolphin-2.6-mistral-7b-dpo-laser
It follows the implementation of laserRMT @ https://github.com/cognitivecomputations/laserRMT and the novel training technique - we partially freeze the model according to a laser-like analysis (Official Paper soon) which effectively prevents the significant problem of language models forgetting previously acquired knowledge. This aspect is particularly crucial when attempting to teach the model specific skills, such as function calling.
We intend to be the first of a family of experimentations being carried out @ Cognitive Computations.
Usage
This model is trained on a variaton of glaive function calling v2 and uses their format for declaring functions and invoking them. Note that like other Dolphin models you should be using the ChatML syntax for representing chats.
Set your system prompt to something along these lines:
You are a helpful assistant with access to the following functions.
Use them if required -
{ "name": "search_wikipedia", "description": "Search for a given keyword or phrase on Wikipedia", "parameters": { "type": "object", "properties": { "query": { "type": "string", "description": "The keyword or phrase to search" } }, "required": [ "query" ] } }
The model will then attempt to invoke the specified functions using the token <functioncall> and some JSON with a name and an arguments parameter.
e.g.:
ASSISTANT: <functioncall> {"name": "search_wikipedia", "arguments": '{"query": "Eiffel Tower"}'} <|endoftext|>
Your code should detect this and stop execution and invoke the function call and inject a response back into the context with the following syntax:
FUNCTION RESPONSE: {"result": "The Eiffel Tower is a wrought-iron lattice... etc"}
You should then prompt the assistant to generate another response.
Note that the model may not always attempt to invoke a function and may respond saying it doesn't have access to the right functionality.