Awesome AI Agents › AI Agent Deployment

xingyaoww/code-act

⭐ 1700 Python repository created 2024-01-13

The project "Executable Code Actions Elicit Better LLM Agents" (CodeAct) is a research initiative and software implementation aimed at improving the performance of large language model (LLM) agents by using executable code as a unified action space. The core idea is to integrate a Python interpreter with LLM agents, allowing them to execute code actions dynamically, revise previous actions, and generate new actions based on observations such as code execution results. This approach enables multi-turn interactions that enhance the agent's ability to perform complex tasks more effectively. The repository includes the official code and resources for the ICML 2024 paper by Xingyao Wang et al. It provides a comprehensive framework called CodeAct, which outperforms traditional action representations like text and JSON by up to 20% in success rate on benchmarks such as API-Bank and the newly curated M3ToolEval. The project also offers a large instruction-tuning dataset named CodeActInstruct, consisting of 7,000 multi-turn interactions, which is publicly available on Hugging Face. Additionally, the project features CodeActAgent, a trained LLM agent based on the CodeAct framework, with two variants using Mistral-7b and Llama-2-7b models. These agents excel in out-of-domain tasks without compromising general conversational abilities. The repository provides detailed instructions for deploying the model using various methods, including Docker with vLLM, llama.cpp for local laptop inference, and Kubernetes for scalable deployment. The system architecture includes components for LLM serving, interaction interfaces (chat UI or Python scripts), and a containerized code execution engine that runs code in isolated Docker containers per chat session. The project also offers a live chat demo and links to the paper, dataset, and models for easy access. Overall, CodeAct represents a significant advancement in LLM agent design by leveraging executable code to unify and enhance agent actions, supported by extensive evaluation, datasets, and deployment tools.

https://github.com/xingyaoww/code-act

api-bankchat-uicode-execution-enginecodeactcodeactagentcodeactinstructdockerexecutable-codeicml-2024instruction-tuning-datasetkubernetesllama-2-7bllama.cppllmllm-agentllm-agentsllm-finetuningllm-frameworkm3toolevalmistral-7bmulti-turn-interactionsopen-source-modelsout-of-domain-taskspython-interpretervllm