eth-sri/smiles-eval

这是一个用于衡量大型语言模型在生成自然语言描述的化学分子SMILES表示方面的数据集。数据集通过提示Gemini 2.5 Pro生成分子描述和对应的SMILES编码，经过筛选有效分子、可靠重生的分子以及去重处理后形成。数据集根据Gemini 2.5 Pro生成分子的可靠性分为不同的难度类别。

This is a dataset for measuring the capabilities of large language models at generating SMILES chemical molecule representations from natural language descriptions. The dataset is created by prompting Gemini 2.5 Pro to generate molecule descriptions and corresponding SMILES codes, which are then filtered for valid molecules, reliably regenerated molecules, and duplicates are removed. The dataset is categorized by different difficulty levels based on the reliability of Gemini 2.5 Pro in generating molecules.