neuripssubmission737/uspto_advanced_v1

USPTO-Advanced是一个基于公共USPTO反应集合的属性分层单步逆合成基准数据集。该数据集通过从USPTO反应集合中筛选，旨在扩展对长尾反应属性轴的覆盖，包括稀有反应类别、大反应中心、螺环/桥环骨架、多组分反应、部分指定立体化学和新颖Bemis–Murcko骨架，同时确保测试集与主流逆合成训练语料库（如USPTO-50k、USPTO-Full、USPTO-MIT、USPTO-STEREO）在产品层面不相交。数据集包含三个测试层：test_2k_hard（挑战子集）、test_5k（匹配分布的标准测试）和test_10k（综合测试），以及一个验证集和一个产品不相交的训练集。数据通过七个逆合成属性轴进行分层采样，包括反应类别、反应中心、骨架拓扑、骨架大小、骨架新颖性、组分数量和立体化学。每个数据行包含反应字符串、标识符和用于分层的特征轴。数据集适用于逆合成模型的训练和评估，支持化学信息学和机器学习研究。USPTO-Advanced is a property-stratified single-step retrosynthesis benchmark derived from the public USPTO reaction collection. Each test row is selected to broaden coverage of the long-tailed reaction-property axes (rare reaction classes, large reaction centers, spiro / bridged scaffolds, multi-component reactions, partially-specified stereo, novel Bemis–Murcko scaffolds) while holding the test split product-disjoint from the dominant retrosynthesis training corpora (USPTO-50k, USPTO-Full / GLN-Full, USPTO-MIT, USPTO-STEREO). The release ships three test tiers — test_2k_hard (challenge subset), test_5k (matched-distribution standard test), and test_10k (comprehensive test) — together with a validation split and a product-disjoint train split. The dataset is stratified along seven retrosynthesis property axes, including reaction class, reaction center, scaffold topology, scaffold size, scaffold novelty, component count, and stereochemistry. Each row contains reaction strings, identifiers, and the feature axes used for stratification, making it suitable for training and evaluating retrosynthesis models in cheminformatics and machine learning.