Deduplicated the prompts
Reformatted the prompts and ground truth answers to be compatible with TRL's GRPO trainer
We have also derived pure English and Chinese subsets.
The full dataset processing logic can be found in create_dataset.py.
If you find this dataset useful in your work, please cite the original source with:
@misc{yu2025dapoopensourcellmreinforcement… See the full description on the dataset page:
https://huggingface.co/datasets/open-r1/DAPO-Math-17k-Processed.
