Anonymous-NeurIPS26-TabularMath/TabularMath

TabularMath是一个程序验证的表格基准测试数据集，旨在诊断表格机器学习模型是否能够从统计插值转向计算外推。数据集包含114个表格回归任务，每个任务由数学文字问题（如小学算术或竞赛数学问题）编译为Python (generator, verifier)对，并针对原始种子答案进行验证，确保标签噪声为零。每个任务有2,048行数据，目标变量y为精确整数。数据集分为两个子集：gsm8k（基于GSM8K小学算术问题，100个任务，204,800行）和aime（基于AIME 2024竞赛数学问题，14个任务，28,672行），总行数为233,472。每个任务的表格包含原始数值槽位（slot_*列）、确定性通用变换列（如绝对值对数变换、奇偶性、模运算等）和整数目标列y。数据集支持多种加载方式，包括使用pandas、huggingface_hub或datasets库，并推荐使用随机分割和OOD排序目标分割进行评估，以测试模型在输出外推下的表现。TabularMath is a program-verified tabular benchmark that probes whether tabular machine-learning models can move from statistical interpolation to computational extrapolation. It consists of 114 tabular regression tasks, each compiled from a math word problem (e.g., grade-school arithmetic or competition math) into a Python (generator, verifier) pair validated against the original seed answer, resulting in zero label noise. Each task has 2,048 rows with exact integer targets (y). The dataset includes two subsets: gsm8k (based on GSM8K grade-school arithmetic, 100 tasks, 204,800 rows) and aime (based on AIME 2024 competition math, 14 tasks, 28,672 rows), totaling 233,472 rows. Each task-specific table contains primitive numeric slots (slot_* columns), deterministic generic transforms (e.g., absolute log1p, evenness, modulo operations), and the integer target column y. It supports multiple loading methods (e.g., pandas, huggingface_hub, datasets) and recommends evaluation protocols with random and OOD sorted-target splits to test output-extrapolation performance.