Bdyskov/verified-math-reasoning
收藏资源简介:
verified-math-reasoning(CargoDash旗舰配方)是一个数学推理数据集,包含999行数据,展示了三层、程序验证、投票分层的数学推理轨迹。该数据集专门设计为小规模,旨在证明CargoDash框架在真实生产LLM端点上的工作效果,而非作为一个严肃的数学基准。每行数据包含针对同一问题的三个独立思维链解决方案,分别来自DeepSeek、Doubao和Qwen3.5模型,每个解决方案都程序化提取了答案并与NuminaMath参考答案进行数值比较。数据根据正确模型数量分为三个子集:train_clean(524行,所有三个模型都答对,用于高置信度的SFT/蒸馏)、hard_disagreement(312行,1-2个模型答对,作为困难基准/RLVR难度数据)和unsolved_flagged(163行,所有模型都答错,需在训练前审查)。数据集基于NuminaMath种子生成,经过去污染处理(从1000个种子中删除了1个污染行),并包含详细的验证和许可信息。
verified-math-reasoning (CargoDash flagship recipe) is a 999-row showcase of three-layer, program-verified, vote-stratified math reasoning traces. The dataset is small on purpose — its job is to prove the CargoDash framework works on real production LLM endpoints, not to be a serious math benchmark. Each row carries three independent chain-of-thought solutions to the same problem (from DeepSeek, Doubao, and Qwen3.5) plus a programmatically extracted answer from each, compared numerically against the NuminaMath reference. Rows are routed by how many models got it right into three splits: train_clean (all 3 models agreed on the correct answer, 524 rows), hard_disagreement (1 or 2 of 3 models got it right, 312 rows), and unsolved_flagged (all 3 failed, 163 rows). The dataset is derived from NuminaMath seeds with decontamination (1 contaminated row dropped from 1000 seeds), and includes detailed verification and licensing information.




