Overview

marclove/llama-2-7b-chat-functions is a 7 billion parameter model, fine-tuned by Marc Love from the base Llama-2-7b-chat-hf model. Its unique training regimen involves a 50/50 mix of synthetic OpenAPI function calls paired with natural language invocations, and chat completions from the Guanaco subset of the OASST1 dataset. This dual-dataset approach aims to balance function calling accuracy with general chat model helpfulness.

Key Capabilities

• Function Calling: Generates structured OpenAPI function calls based on natural language prompts.
• Chat Completions: Retains general conversational abilities for non-function-calling scenarios.
• Beta State: Currently in beta, with future updates planned to improve performance and potentially modify the prompting format.

Good For

• Research and Development: Ideal for exploring and experimenting with function calling capabilities in smaller LLMs.
• Prototyping: Useful for quickly prototyping applications that require models to interact with external tools via function calls.
• Validation: Users are encouraged to validate all function call outputs due to potential hallucinations, especially in the 7B model. Tools like Pydantic are suggested for validation.