MolLangData: A Large-Scale Dataset of Paired Molecular Structures and Natural Language Descriptions

MolLangData is a large-scale dataset containing 163,111 paired samples of molecular structures and natural language descriptions, generated via a rule-regularized method using large language models. The dataset comprises two subsets: generated_data (161,111 rows) containing AI-generated descriptions across easy, medium, and hard difficulty levels; and validated_data (2,000 rows) containing curated and human-validated examples with 98.6% overall precision. Each sample includes the compound's CID, SMILES notation, IUPAC name, and a natural language description (375–8,010 characters). The dataset is intended for training and evaluating molecular language models, supporting tasks such as molecular structure recognition, description generation, and molecular property reasoning.