EnvScaler-SFT-Traj-9K

Dataset Description

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

This dataset contains 9,022 supervised fine-tuning (SFT) trajectories generated by the EnvScaler framework. Each trajectory records the complete interaction history between an LLM agent and a tool-interactive environment, including the conversation messages, tool calls, and intermediate steps.

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

Dataset Structure

Each trajectory contains the following fields:

• traj_type: Type of trajectory, either "conversation" or "non_conversation", indicating whether the trajectory includes user messages
• task_info: Dictionary containing task information:
  • env_id: Unique identifier for the environment (e.g., env_1_sft)
  • task_id: Unique identifier for the task/scenario (e.g., env_1_sft-task_1)
  • task: Task description and instructions
• tools: Tool interface definitions in JSON format (stored as JSON string, use json.loads() to convert to a list). Contains the available tools/functions the agent can use
• messages: Complete conversation messages in JSON format (stored as JSON string, use json.loads() to convert to a list). Contains the full dialogue history including system, user, assistant, and tool messages
• user_messages: User messages in JSON format (stored as JSON string, use json.loads() to convert to a list). Contains only the user's messages from the conversation
• steps: Number of steps in the trajectory, representing the number of interactions with the environment

Dataset Statistics

• Number of trajectories: 9,022
• Language: English
• Format: JSON

Data Source

This dataset is automatically generated by collecting trajectories from agent-environment interactions using the EnvScaler framework. Trajectories are collected by having Qwen3-30B-A3B-Thinking-2507 interact with the synthesized environments from the SFT scenarios.

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]