strablebench/STRABLE

STRABLE是一个包含108个真实世界表格数据集的基准语料库，这些数据集含有字符串特征，旨在支持对处理字符串条目的表格机器学习管道进行实证研究。该数据集填补了现有表格基准的空白，后者通常排除字符串列或在评估前将其扁平化为固定数值表示。STRABLE提供了原始、经过最小预处理的表格，涵盖8个应用领域（经济、健康、基础设施、能源、教育、商业、食品、社会），包括13个二分类任务、19个多类分类任务和76个回归任务。关键统计包括：来自33个不同公共来源，中位数为7,700行、18列、每个单元格17个字符、每个字符串列1,200个唯一值；字符串分类包括分类（50%）、名称（23%）、结构化代码（17%）、自由文本（8%）和标识符（0.5%）。数据集应用了最小预处理，以反映实践中的数据状态，例如扁平化嵌套结构、删除重复行和单值列，但不进行特征工程或缺失值插补。

STRABLE is a benchmarking corpus of 108 real-world tabular datasets containing string features, designed to support empirical research on tabular machine learning pipelines that handle string entries. It fills the gap where existing tabular benchmarks either exclude string columns or flatten them into fixed numerical representations before evaluation. STRABLE provides raw, minimally preprocessed tables spanning 8 application fields (Economy, Health, Infrastructure, Energy, Education, Commerce, Food, Social), covering 13 binary classification, 19 multi-class classification, and 76 regression tasks. Key statistics include: sourced from 33 distinct public sources; median of 7,700 rows, 18 columns, 17-character strings per cell, and 1,200 unique values per string column; string taxonomy includes Categorical (50%), Names (23%), Structured Codes (17%), Free Text (8%), and Identifiers (0.5%). Minimal preprocessing is applied to reflect data as found in practice, such as flattening nested structures and removing duplicate rows and single-value columns, without feature engineering or missing values imputation.