EnvScaler-RL-Scenario

Dataset Description

Usage Note: The init_config field is stored as a JSON string. You need to use json.loads() to convert it to a dictionary when using the dataset.

This dataset contains 2,550 reinforcement learning (RL) scenarios generated by the EnvScaler framework. Each scenario consists of an environment configuration, a task description, and a checklist with check functions for reward calculation in reinforcement learning training.

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

Dataset Structure

Each scenario contains the following fields:

• env_id: Unique identifier for the environment (e.g., env_148_rl)
• env_class_name: Python class name of the environment (e.g., ContactLensSubscriptionManagementSystem)
• task_id: Unique identifier for the task/scenario (e.g., env_148_rl-task_1)
• init_config: Initial state configuration in JSON format (stored as JSON string, use json.loads() to convert to a dictionary). This configuration can be used to initialize the environment state
• task: Task description and instructions specifying what the agent needs to accomplish in the environment
• checklist_with_func: List of checklist items with corresponding Python check functions. Each item contains:
  • check_item: Description of what needs to be checked
  • check_func: Python function code (as a string) that evaluates whether the check item is satisfied based on the final state of the environment. These functions are used for reward calculation in RL training

Dataset Statistics

• Number of scenarios: 2,550
• Language: English
• Format: JSON

Data Source

This dataset is automatically generated by the ScenGenerator module of the EnvScaler framework. ScenGenerator synthesizes multiple scenarios for each environment skeleton through the following steps:

1. Generate an initial state/database configuration for the environment
2. Create a challenging task that can be solved from that state
3. Decompose the task into checklists and convert each checkpoint into a Python Boolean function over the final state of the environment

The checklist functions provide rule-based, verifiable reward signals for reinforcement learning, enabling precise evaluation of agent performance.

Related Resources

• Project Homepage: EnvScaler GitHub
• Paper: Arxiv
• Other Datasets:
  • EnvScaler-191-Env
  • 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]