Name: PropMolFlow_QM9_CNOFH_2025 Extended ID: PropMolFlow_QM9_CNOFH_2025__Zeng-Jin-Karypis-Transtrum-Tadmor-Hennig-Roitberg-Martiniani-Liu__DS_6qqf55wad1mv_0 Description: This DFT dataset is curated in response to the growing interest in property-guided molecule genaration using generative AI models. Typically, the properties of generated molecules are evaluated using machine learning (ML) property predictors trained on fully relaxed dataset. However, since generated molecules may deviate significantly from relaxed structures, these predictors can be highly unreliable for assessing their quality. This data provides DFT-evaluated properties, energy and forces for generated molecules. These structures are unrelaxed and can serve as a validation set for machine learning property predictors used in conditional molecule generation. It includes 10,773 molecules generated using PropMolFlow, a state-of-the-art conditional molecule generation model. PropMolFlow employs a flow matching process parameterized with an SE(3)-equivariant graph neural network. PropMolFlow models are trained on QM9 dataset. Molecules are generated by conditioning on six properties---polarizibility, gap, HOMO, LUMO, dipole moment and heat capacity at room temperature 298K---across two tasks: in-distribution and out-of-distribution generation. Full details are available in the corresponding paper. Authors: Cheng Zeng Jirui Jin George Karypis Mark Transtrum Ellad B. Tadmor Richard G. Hennig Adrian Roitberg Stefano Martiniani Mingjie Liu DOI: 10.60732/1f7cae3c Calculated Property Types: atomic_forces energy Elements: C (34.87%) F (0.18%) H (54.17%) N (4.48%) O (6.3%) Methods: DFT-B3LYP Software: Gaussian 16 Number of Configurations: 10,773 Number of Atoms: 205,304 Publication Link: https://arxiv.org/abs/2505.21469