CodeActAgent-Llama-2-7b: Executable Code Actions for LLM Agents
CodeActAgent-Llama-2-7b is a 7 billion parameter model built on the Llama-2 architecture, developed by Xingyao Wang and collaborators. It introduces the CodeAct paradigm, which unifies LLM agent actions into executable Python code. This approach allows agents to interact with a Python interpreter, dynamically revising actions or generating new ones based on code execution results and observations in multi-turn interactions.
Key Capabilities & Features
- Executable Code Actions: Utilizes Python code as a unified action space, enabling more robust and dynamic agent behavior.
- Enhanced Agent Performance: Outperforms traditional text and JSON action formats, achieving up to 20% higher success rates on benchmarks like API-Bank and M3ToolEval.
- Multi-turn Interaction: Designed for iterative problem-solving, allowing agents to adapt and refine their actions based on real-time feedback from code execution.
- CodeActInstruct Dataset: Trained on a specialized dataset of 7k multi-turn interactions, focusing on code-based actions.
- Out-of-Domain Task Excellence: Excels in agent tasks beyond its immediate training domain, while maintaining strong general conversational and knowledge performance.
Why Choose CodeActAgent-Llama-2-7b?
This model is particularly suited for applications requiring LLM agents to perform complex, verifiable actions through code. Its ability to execute and revise code makes it a strong candidate for automated problem-solving, tool use, and interactive environments where precise and dynamic control is crucial. While CodeActAgent-Mistral-7b-v0.1 is recommended for its larger context window, this Llama-2 variant offers a capable alternative for scenarios where a 4k context window is sufficient.