introvoyz041/completeRXN

CompleteRxn Benchmark是一个包含206,423个反应的化学数据集，专为反应完成任务设计。给定一个原子不平衡的USPTO反应SMILES（简化分子输入线输入系统），目标是预测缺失的分子以生成平衡的化学反应方程式。真实反应目标来源于FlowER机制步骤，这些步骤已映射到USPTO记录。数据集提供三种分割类型（随机分割、组外分布分割和极端外分布分割），每种类型有5次重复，以测试模型在不同结构新颖性水平下的泛化能力。数据集还包括元数据，如反应标识符、原始反应、输入反应、目标反应、立体化学信息、原子平衡状态、反应类别和专利信息等，适用于化学信息学和机器学习基准测试。

The CompleteRxn Benchmark is a dataset of 206,423 reactions for the task of reaction completion: given an atom-unbalanced USPTO reaction SMILES, predict the missing molecules to produce a balanced equation. Ground-truth targets come from FlowER mechanistic steps mapped to USPTO records. It includes three split types (random, group OOD, extreme OOD) with 5 repetitions each to test generalization across structural novelty levels. The dataset features metadata such as unique identifiers, original reactions, input reactions, target reactions, stereochemistry details, atom balance status, reaction classes, and patent information, suitable for cheminformatics and machine learning benchmarking.