Anonymous-07/SafeChem
收藏资源简介:
SafeChem是一个基于法规的基准数据集,包含32,614种化学物质,用于多标签GHS(全球统一制度)危害预测和LLM安全可靠性评估。与之前通过查询药物数据库构建的分子基准不同,SafeChem源自一个经过筛选的有害物质注册表,确保覆盖现实世界工业和关键安全化学品,包括溶剂、反应性气体、农用化学品和重金属化合物,这些在药物为中心的基准中代表性不足。每种物质都标注有分子结构表示(规范SMILES、InChI、InChI键)、八个物理化学描述符、自由文本物理描述以及30个二元GHS危害分类标签,涵盖毒理学、物理和环境危害类别,这些标签源自欧洲法规来源(CLP法规和ECHA C&L库存,由PubChem汇总)。数据集支持两个基准任务:1)多标签危害预测:从分子结构对17个保留的危害类别进行支架分割评估;2)LLM安全幻觉基准:在500种高危害物质子集上评估LLM的遗漏幻觉率、委托幻觉和静默弃权行为。
SafeChem is a regulatory-grounded benchmark dataset of 32,614 chemical substances for multi-label GHS (Globally Harmonized System) hazard prediction and LLM safety reliability evaluation. Unlike prior molecular benchmarks constructed by querying pharmaceutical databases, SafeChem is seeded from a curated hazardous materials registry, ensuring coverage of real-world industrial and safety-critical chemicals including solvents, reactive gases, agrochemicals, and heavy metal compounds underrepresented in drug-focused benchmarks. Each substance is annotated with molecular structure representations (canonical SMILES, InChI, InChI key), eight physicochemical descriptors, free-text physical descriptions, and 30 binary GHS hazard classification labels spanning toxicological, physical, and environmental hazard categories derived from European regulatory sources (CLP Regulation and ECHA C&L Inventory as aggregated by PubChem). The dataset supports two benchmark tasks: 1) Multi-label hazard prediction: Scaffold-split evaluation of GHS hazard classification from molecular structure across 17 retained hazard classes with sufficient test-set support; 2) LLM safety hallucination benchmark: Evaluation of omission hallucination rate, commission hallucination, and silent abstention behavior across LLMs on a curated 500-substance high-hazard subset.




