Affine-5647568

AgentEvol-7B

<p align="center"> 📃 <a href="TODO" target="_blank">Paper</a > • 🌐 <a href="https://agentgym.github.io/" target="_blank">Project Page</a > • 💻 <a href="https://github.com/WooooDyy/AgentGym" target="_blank">[Github Repo]</a> • 📚 <a href="https://huggingface.co/datasets/AgentGym/AgentTraj-L" target="_blank">[Trajectory Dataset]</a > • 📈 <a href="https://huggingface.co/datasets/AgentGym/AgentEval" target="_blank">[Eval Benchmark]</a> • 🤗 <a href="https://huggingface.co/AgentGym/AgentEvol-7B" target="_blank">Model (AgentEvol-7B)</a ><br> </p >

AgentEvol is a novel method to evolve generall-capable LLM-based agents across multiple environments. AgentEvol first trains a base generally-capable agent with behavioral cloning, equipping it with basic abability and prior knowledgs. Subsequently, the agent is allowed to perform exploration and learning acorss various tasks and environments.

AgentEvol-7B is trained with the AgentEvol algorithm on Llama-2-Chat-7B. The model is first trained on the AgentTraj set with behavioural cloning. Next it performs exploration and learning from a broader set of instructions. After evolution, its performance outperforms SOTA models on many tasks.

🔖 Citation

@misc{xi2024agentgym,
      title={AgentGym: Evolving Large Language Model-based Agents across Diverse Environments}, 
      author={Zhiheng Xi and Yiwen Ding and Wenxiang Chen and Boyang Hong and Honglin Guo and Junzhe Wang and Dingwen Yang and Chenyang Liao and Xin Guo and Wei He and Songyang Gao and Lu Chen and Rui Zheng and Yicheng Zou and Tao Gui and Qi Zhang and Xipeng Qiu and Xuanjing Huang and Zuxuan Wu and Yu-Gang Jiang},
      year={2024},
      eprint={2406.04151},
      archivePrefix={arXiv},
      primaryClass={cs.AI}
}