ScaProp-Instructions: A Multimodal Large Language Model scaffold Constrained Molecular Generation Dataset

Based on these property changes, together with predefined prompt templates, natural-language instructions are reconstructed to describe the desired property constraints for molecular generation.The resulting samples are organized as multimodal triplets consisting of: (1) a molecular scaffold image, (2) a natural-language property instruction, and (3) a target molecular structure represented by SMILES. These triplets provide aligned visual, textual, and molecular information for training and evaluating multimodal large language models in scaffold-preserving molecular generation.The dataset is subsequently divided into training, validation, and test sets according to a predefined ratio.

The test set is constructed to contain independent molecular samples with non-overlapping molecular skeleton topologies relative to the training and validation sets, enabling evaluation of model generalization to unseen chemical scaffolds.ScaProp-Instructions is intended to support research on multimodal molecular generation, scaffold preservation, property-controlled molecular optimization, and vision-language alignment in the chemical domain.