Please refer to the 📺GitHub repo and 📃Paper for more details.
This dataset contains 301,909 single-step (Proof State, DExploration Action) pairs extracted from 39,509 DExploration trajectories. It is produced by applying exploratory transformation on the NuminaMath-Lean dataset: formal statement-proof pairs are transformed into mathematical exploration trajectories through deductive rewriting, dependency analysis, and exploratory reassembling.
Each sample represents one step in the DExploration process, where the model is given the current exploration state and must propose the next action:
Each line in exploration_steps.jsonl is a JSON object representing a single-turn conversation:
{
"conversation": [
{
"system": "<system prompt>",
"input": "<user prompt with exploration state>",
"output": "<model response with next step>"
}
]
}
| Field | Description |
|---|---|
system | System prompt (role description) |
input | User input (task description + proof state) |
output | Model output (DExploration action) |
If you find our work useful in your research, please cite:
@inproceedings{
liu2026lets,
title={Let's Explore Step by Step: Generating Provable Formal Statements with Deductive Exploration},
author={Qi Liu and Kangjie Bao and Yue Yang and Xinhao Zheng and Renqiu Xia and Qinxiang Cao and Junchi Yan},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=Njrkeo3DiJ}
}
This project is released under the Apache 2.0 license. See LICENSE for details.
We welcome contributions! Please feel free to submit issues or pull requests.
For questions about the paper, data, or code:
Please refer to the 📺GitHub repo and 📃Paper for more details.
This dataset contains 301,909 single-step (Proof State, DExploration Action) pairs extracted from 39,509 DExploration trajectories. It is produced by applying exploratory transformation on the NuminaMath-Lean dataset: formal statement-proof pairs are transformed into mathematical exploration trajectories through deductive rewriting, dependency analysis, and exploratory reassembling.
Each sample represents one step in the DExploration process, where the model is given the current exploration state and must propose the next action:
Each line in exploration_steps.jsonl is a JSON object representing a single-turn conversation:
{
"conversation": [
{
"system": "<system prompt>",
"input": "<user prompt with exploration state>",
"output": "<model response with next step>"
}
]
}
| Field | Description |
|---|---|
system | System prompt (role description) |
input | User input (task description + proof state) |
output | Model output (DExploration action) |
If you find our work useful in your research, please cite:
@inproceedings{
liu2026lets,
title={Let's Explore Step by Step: Generating Provable Formal Statements with Deductive Exploration},
author={Qi Liu and Kangjie Bao and Yue Yang and Xinhao Zheng and Renqiu Xia and Qinxiang Cao and Junchi Yan},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=Njrkeo3DiJ}
}
This project is released under the Apache 2.0 license. See LICENSE for details.
We welcome contributions! Please feel free to submit issues or pull requests.
For questions about the paper, data, or code: