AceMath RewardBench is a dataset released by NVIDIA in 2024 for evaluating the capabilities of mathematical reward models, with the related research paper presented in "AceMath: Advancing Frontier Math Reasoning with Post-Training and Reward Modeling", aiming to comprehensively measure the performance of mathematical reasoning reward models through a best-of-N (N=8) setting.

The dataset includes samples from 7 datasets: GSM8K (1319 questions), Math500 (500 questions), Minerva Math (272 questions), Gaokao 2023 en (385 questions), OlympiadBench (675 questions), College Math (2818 questions), and MMLU STEM (3018 questions). Each sample contains a mathematical problem, 64 answer attempts of varying quality generated by 8 different language models such as Qwen2/2.5-Math, LLama3.1, Mathtral, and deepseek-math, ground truth scores for each answer, and metadata such as problem difficulty and topic domain. This benchmark focuses on diversity and robustness, evaluating average results through random sampling with 100 random seeds.

Dataset Composition

The dataset contains the following fields:

• question: The textual description of the mathematical problem
• code: A list of complete answers/solutions generated by the model
• gt: The ground truth answer
• pred: A list of predicted results extracted from each model response
• score: A boolean list indicating whether each response matches the ground truth answer
• report: Related report
• idx: Index
• gt_cot: Ground truth chain of thought

Citation

@article{acemath2024,
  title={AceMath: Advancing Frontier Math Reasoning with Post-Training and Reward Modeling},
  author={Liu, Zihan and Chen, Yang and Shoeybi, Mohammad and Catanzaro, Bryan and Ping, Wei},
  journal={arXiv preprint},
  year={2024}
}