sungyub/math-verl-unified

Math-VERL Unified Collection是一个统一的数学推理数据集集合，包含9个数据集，总计226万9166个去重问题，所有问题都已转换为VERL格式，用于强化学习应用。这个集合结合了来自竞赛级问题到高级推理任务的多样化数学内容。数据集的特点包括：2.27M个去重样本，跨数据集去重，统一的VERL格式，难度级别多样，涵盖多个数学领域，优先级去重。数据集源自9个精心挑选的数据源，包括数学竞赛、高级推理任务等。预处理流程包括数据集特定清洗和全局去重。VERL格式是一个标准化的格式，用于验证和强化学习。数据集统计信息包括样本分布、数学领域覆盖和质量指标。使用说明提供了如何加载整个集合、特定源数据集和使用VERL的示例。数据集的引用信息包括统一集合和原始源数据集的引用格式。数据集采用混合许可证，每个源数据集保留其原始许可证。致谢部分感谢了原始数据集作者、处理工具和开源机器学习社区。版本历史记录了数据集的更新和改进。相关资源包括单个数据集、文档和其他数学集合。

The Math-VERL Unified Collection is a unified collection of mathematical reasoning datasets, including 9 datasets with a total of 2,269,166 deduplicated problems, all converted to the VERL format for reinforcement learning applications. The collection combines diverse mathematical content from competition-level problems to advanced reasoning tasks. Key features include 2.27M deduplicated samples, inter-dataset deduplication, unified VERL format, diverse difficulty levels, multiple mathematical domains, and priority-based deduplication. The collection aggregates 9 datasets from various sources, each processed through a tailored preprocessing pipeline. The VERL schema is a standardized format for verification and reinforcement learning. Dataset statistics include split distribution, mathematical domain coverage, and quality metrics. Usage examples demonstrate how to load the full collection, specific source datasets, and use the dataset with VERL. Citation information includes the unified collection and original source datasets. The dataset uses mixed licenses, retaining the original license of each source dataset. Acknowledgments thank the original dataset authors, processing tools, and the open-source ML community. Version history records updates and improvements to the dataset. Related resources include individual datasets, documentation, and other math collections.