Hugging Face Octopus V4: Graph of Language Models
Octopus V4, developed by NexaAIDev, is a 4 billion parameter open-source language model functioning as a master node in a system of specialized LLMs. Its primary role is to efficiently direct user queries to the most suitable domain-specific model, particularly excelling in topics covered by the MMLU benchmark.
Key Capabilities
- Query Routing: Accurately maps user queries to specialized models using a functional token design.
- Query Reformatting: Converts natural language queries into more professional and precise formats for improved response accuracy.
- Compact Size: Designed for efficient and swift operation, including on smart devices.
- MMLU Performance: Achieves a 74.8% MMLU score, outperforming models like GPT-3.5 (70.0%) and Llama3-8b-instruct (68.4%) in a 5-shot learning setup.
Use Cases
Octopus V4 is ideal for applications requiring precise query handling and routing to specialized AI agents. For example, it can direct a query like "Tell me the result of derivative of x^3 when x is 2?" to a dedicated math GPT, reformatting it for optimal processing. The model leverages a selection of domain-specific LLMs for various categories, including biology, physics, computer science, math, health, and law, with a Domain LLM Leaderboard available for exploration.