EnvScaler-191-Env

Dataset Description

Usage Note: The env_func_details and tools fields are stored as JSON strings. You need to use json.loads() to convert them to lists when using the dataset.

This dataset contains metadata for 191 executable, stateful, tool-interactive environments generated by the EnvScaler framework.

EnvScaler is an automated, scalable framework that realizes executable, stateful, tool-interactive environments via programmatic synthesis, for training Large Language Model (LLM) agents.

Dataset Structure

Each environment metadata contains the following fields:

• env_id: Unique identifier for the environment (e.g., env_1_sft)
• environment_summary: Brief summary of the environment (e.g., Task tracking system)
• environment_introduction: Detailed introduction of the environment, describing its purpose, functionality, and use cases
• state_space_definition: State space definition containing multiple entities, each entity contains:
  • entity: Entity name
  • attributes: List of attributes
  • description: Entity description
• constraints_rules: List of constraint rules defining behavioral constraints and business rules of the environment
• operation_list: List of operations, containing two types of operations:
  • query: Query operations
  • state_change: State change operations Each operation contains operation_name, operation_description, and operation_type
• env_class_name: Python class name of the environment class
• env_class_code: Complete Python code of the environment class that can be executed directly
• env_class_def: Definition part of the environment class
• env_func_details: Detailed description of environment functions
• tools: Tool interface definition in JSON format, can be used for Function Calling

Dataset Statistics

• Number of environments: 191
• Language: English
• Format: JSON

Data Source

This dataset is automatically generated by the SkelBuilder module of the EnvScaler framework. SkelBuilder builds environments through the following steps:

1. Mine potential environment descriptions from existing open-source textual tasks
2. Plan corresponding state schema, business rules, and tools
3. Generate fully functional Python classes whose methods expose tool interfaces
4. Perform a dual-agent loop for environment quality inspection (one agent invokes tools, the other checks code, return values, and state changes)

Related Resources

• Project Homepage: EnvScaler GitHub
• Paper: Arxiv
• Other Datasets:
  • EnvScaler-SFT-Scenario
  • EnvScaler-RL-Scenario
  • EnvScaler-SFT-Traj-9K
• Models:
  • EnvScaler-Qwen3-1.7B
  • EnvScaler-Qwen3-4B
  • EnvScaler-Qwen3-8B

Citation

If you use this dataset, please cite our work:

@article{song2026envscaler,
  title={EnvScaler: Scaling Tool-Interactive Environments for LLM Agent via Programmatic Synthesis},
  author={Song, Xiaoshuai and Chang, Haofei and Dong, Guanting and Zhu, Yutao and Dou, Zhicheng and Wen, Ji-Rong},
  journal={arXiv preprint arXiv:2601.05808},
  year={2026}
}

License

This dataset is licensed under the MIT License.

Contact

For any questions or feedback, please contact: [email protected]