metadata
license: mit
task_categories:
- text-generation
language:
- en
tags:
- diffusion-models
- reinforcement-learning
- math-reasoning
- code-generation
- reasoning
Revolutionizing Reinforcement Learning Framework for Diffusion Large Language Models Datasets
This repository contains datasets used in the paper Revolutionizing Reinforcement Learning Framework for Diffusion Large Language Models. These datasets are crucial for building, training, and deploying Diffusion Large Language Models (DLMs) within the TraceRL framework, particularly for improving reasoning performance on complex math and coding tasks.
- Paper: Revolutionizing Reinforcement Learning Framework for Diffusion Large Language Models
- Code (GitHub): https://github.com/Gen-Verse/dLLM-RL
- Project Page (Hugging Face Collection): https://huggingface.co/collections/Gen-Verse/trado-series-68beb6cd6a26c27cde9fe3af
Sample Usage (Data Download)
You can navigate to the ./data directory within the associated GitHub repository to download datasets for evaluation and training. In that directory, you will also find detailed instructions on how to modify your own dataset.
For example, to download the MATH500 and MATH_train datasets:
cd data
python download_data.py --dataset MATH500
python download_data.py --dataset MATH_train
cd ..
Citation
If you use these datasets in your research, please cite the associated paper:
@article{wang2025trado,
title={Revolutionizing Reinforcement Learning Framework for Diffusion Large Language Models},
author={Wang, Yinjie and Yang, Ling and Li, Bowen and Tian, Ye and Shen, Ke and Wang, Mengdi},
journal={arXiv preprint arXiv:2509.06949},
year={2025}
}