Hard2Verify
收藏资源简介:
Hard2Verify是一个人类标注的、基于步骤的验证基准数据集,用于评估前沿数学问题的验证器性能。该数据集包含来自最近国际数学竞赛(如IMO和Putnam)的难题,并使用三个顶级LLM(GPT-5(高)、Gemini 2.5Pro和Claude Sonnet 4(思考))对问题进行回答。最后,由博士级别的数学专家对每个模型生成的步骤进行标注。该数据集的创建历时超过500小时的人力工作,并经过了三轮独立的一致性检查。Hard2Verify旨在解决开放性问题,要求验证器不仅要评估步骤的正确性,还要评估所引用结果的正确性和适用性。该数据集适用于自然发生的应用场景,验证器必须评估模型生成的回答,这些回答通常与人类编写的参考答案有很大的不同。
Hard2Verify is a human-annotated step-by-step verification benchmark dataset designed to evaluate the performance of state-of-the-art mathematical problem verifiers. The dataset comprises challenging problems sourced from recent international mathematics competitions such as the International Mathematical Olympiad (IMO) and the Putnam Mathematical Competition, with responses generated by three top-tier large language models (LLMs): GPT-5 (High), Gemini 2.5 Pro, and Claude Sonnet 4 (Thinking). Each step of the model-generated responses was subsequently annotated by PhD-level mathematics experts. The curation of this dataset required over 500 hours of manual human labor and underwent three independent consistency checks. Hard2Verify targets open-ended mathematical problems, requiring verifiers to not only evaluate the correctness of individual reasoning steps but also to assess the correctness and applicability of the cited results. This dataset is tailored for real-world application scenarios, where verifiers must assess model-generated responses that often differ significantly from human-written reference solutions.
Hard2Verify数据集概述
数据集基本信息
- 名称: Hard2Verify
- 许可证: CC-BY-NC-4.0
- 任务类别: 文本生成
- 语言: 英语
- 标签: 数学
- 规模类别: n<1K(少于1000个样本)
数据集描述
Hard2Verify是一个人工标注的步骤级验证基准数据集,专门用于评估前沿数学问题中的步骤级验证器。该数据集通过500多小时的人工劳动创建,要求验证器为前沿大语言模型生成的数学问题回答提供步骤级标注或识别第一个错误。
数据集内容
数据格式
解密后的数据集每行包含以下字段:
unique_id: 包含来源奥林匹克竞赛和生成模型信息的字符串question: 原始数学问题model_response_by_step: 按步骤分解的模型解答human_labels: 对应每个步骤的人工正确性标签(0=错误,1=正确)human_labels_first_error_idx: 第一个错误的索引(0索引);-1表示无错误
解密方法
数据集行经过加密处理,需要使用GitHub仓库中的decrypt_sample函数进行解密。
相关资源
- 论文: https://arxiv.org/abs/2510.13744
- 评估代码: https://github.com/SalesforceAIResearch/Hard2Verify
使用限制
- 该数据集使用GPT、Gemini和Claude生成,不得用于开发竞争产品
- 仅限研究用途,支持学术论文
引用信息
bibtex @misc{pandit2025hard, title={Hard2Verify: A Step-Level Verification Benchmark for Open-Ended Frontier Math}, author={Pandit, Shrey and Xu, Austin and Nguyen, Xuan-Phi and Ming, Yifei and Xiong, Caiming and Joty, Shafiq}, year={2025}, journal={arXiv preprint arXiv:2510.13744}, }




