✨ AgentGen: Enhancing Planning Abilities for Large Language Model based Agent via Environment and Task Generation

📢 Updates

• 2025/05: Checking our new Text2World paper on PDDL generation accepted by ACL 2025 Findings.
• 2024/12/11: Initial release the code

🛠️ Installation

1. Configure the environment:

conda env create -f environment.yml
conda activate agentgen

2. Set up OpenAI credentials:

   • Add your OPENAI_API_KEY to src/key.txt

3. Install required dependencies (pddlgym and pddlgym_planners):

pip install pddlgym
# for mac:
# brew install coreutils
git clone https://github.com/ronuchit/pddlgym_planners.git
cd pddlgym_planners
pip install -e .

🚀 Quick Start

Generate the environment and domain for the first iteration:

cd src
bash run.sh

🧪 Experiments

This repository provides our implementation to reproduce the data generation process described in the original AgentGen paper. To replicate the results:

1. Segment the src/data/inpiration_corpus/lima.json
2. Run the run.sh script iteratively

For additional components:

• Model Training: We use llama-factory
• Model Evaluation: We use AgentBoard

📦 Model & Dataset

• Dataset: The dataset after iterative environment and task generation is available in the src/data/it12 folder. The trainig-ready conversation SFT data is available at src/data/sft_data.json.
• Model: We followed the training configuration from the original paper to reproduce the models (We trained the AgentGen-8B model with alpaca template since there exists a bug when training with llama-3 template):

Model	Huggingface Repo
AgentGen-70B-Lora-Rank1	🤗 Huggingface
AgentGen-70B-Lora-Rank16	🤗 Huggingface
AgentGen-8B	🤗 Huggingface

🙏 Cite

If you find this repo helpful in your research, please cite the original paper:

@article{hu2024agentgen,
  title={Agentgen: Enhancing planning abilities for large language model based agent via environment and task generation},
  author={Hu, Mengkang and Zhao, Pu and Xu, Can and Sun, Qingfeng and Lou, Jianguang and Lin, Qingwei and Luo, Ping and Rajmohan, Saravan and Zhang, Dongmei},
  journal={arXiv preprint arXiv:2408.00764},
  year={2024}
}