anshy/Eurus-2-RL-Data-filtered

Eurus-2-RL-Data是一个高质量的强化学习训练数据集，专注于数学和编程问题，并配备结果验证器（数学答案以LaTeX格式提供，编程问题通过测试用例验证）。数学问题来源于NuminaMath-CoT数据集，涵盖从中国高中数学到国际数学奥林匹克竞赛的题目；编程问题来源于APPS、CodeContests、TACO和Codeforces数据集，主要为编程竞赛级别。数据集经过详细清理和过滤：数学方面，使用高级推理模型（如Qwen-QwQ）过滤掉不可解、不匹配或答案错误的问题，并将选择题转换为开放性问题；编程方面，主要过滤重复问题。最终，数据集包含455,261个数学问题和25,276个编程问题，用于训练集，验证集各包含1,024个问题。

Eurus-2-RL-Data is a high-quality RL training dataset of mathematics and coding problems with outcome verifiers (LaTeX answers for math and test cases for coding). For math, it sources from NuminaMath-CoT, spanning from Chinese high school mathematics to International Mathematical Olympiad competition questions. For coding, it sources from APPS, CodeContests, TACO, and Codeforces, mainly programming competition level. The dataset undergoes detailed cleaning and filtering: for math, advanced reasoning models like Qwen-QwQ are used to filter out unsolvable, unmatchable, or incorrect answer problems, and multiple-choice questions are reformatted to open questions; for coding, duplicated problems are filtered out. It retains 455k math problems and 26k coding problems, with training set containing 455,261 math and 25,276 coding problems, and validation set containing 1,024 each.