Datasets
OpenR utilizes multiple datasets on math problems.
Table of contents
🧮 MATH
OpenR use MATH dataset [Hendrycks et al., 2021] for data augmentation.
OpenR augments the data by automatically generating synthetic samples using automated methods such as OmegaPRM. This approach reduces the reliance on expensive human annotations, enabling more scalable data collection. The data-processing code can be found at openr/data.
🎲 MATH-APS
MATH-APS is the MATH dataset processed by OmegaPRM.
We have released MATH-APS on huggingface. For more details, please refer to our paper.
🔢 PRM800K
The original dataset can be found in this GitHub repo.
🧩 Math-Shepherd
The original dataset can be found in this Hugging Face dataset repo. It was used to train the Math-Shepherd PRM.
Integrated Datasets for PRM Training
The PRM is trained through supervised fine-tuning on an LLM, with the correct/incorrect distinction serving as the classification label. We train a PRM named Math-psa using datasets such as PRM800K, Math-Shepherd, and our MATH-APS dataset. For detailed integration please refer to openr/prm/code.