Nexusflow/NexusRaven-13B

NexusRaven-13B: State-of-the-Art Function Calling

NexusRaven-13B is an open-source, commercially viable 13 billion parameter language model developed by Nexusflow, fine-tuned from CodeLlama-13b-Instruct-hf. It is specifically designed to excel in function calling capabilities, surpassing other open-source LLMs of similar sizes and demonstrating competitive performance against larger proprietary models like GPT-4 in specific contexts.

Key Capabilities & Performance:

• Superior Function Calling: Achieves a 95% success rate in using cybersecurity tools (e.g., CVE/CPE Search, VirusTotal) with a demonstration retrieval system, compared to GPT-4's 64% in the same setup.
• Generalization to Unseen Tools: Demonstrates strong zero-shot generalization, performing comparably to GPT-3.5 on tools not encountered during training.
• Commercially Permissive: Trained without proprietary LLM data, offering full control for commercial deployments.
• Python Function Compatibility: Accepts Python function signatures and docstrings to generate function calls, highly compatible with frameworks like LangChain.

Usage & Limitations:

• Optimized for Function Calls: Best used for scenarios requiring the model to select and execute predefined functions based on user queries.
• Stop Criteria Recommended: It is highly recommended to use ["\nReflection:"] as a stop criterion during inference to optimize token usage, as the model's reflection step often does not improve the initial call.
• Context Window: May be limited by its context window when dealing with a very large number of functions, suggesting integration with a retriever for such cases.
• Guardrails: Users should implement guardrails to manage potential incorrect function calls.