FD25

FD25: A Large-Scale Quantum Chemical Dataset for Molecular Property Prediction FD25 is a comprehensive dataset of over 2.6 million organic molecules, computed at the M06-2X/cc-pVDZ level of theory, providing consistent and high-quality thermochemical, electronic, and solvation properties. Designed to support machine learning research in quantum chemistry, FD25 includes total and binding energies, HOMO/LUMO levels, isotropic polarizabilities, Mulliken charges, and implicit solvation energies—all aligned to a single computational protocol to minimize methodological noise. Accompanied by pre-trained graph neural network (GNN) weights and a linear alignment coefficients, FD25 enables accurate property prediction not only within its own domain but also across external datasets (e.g., QM9, PC3M) through systematic bias correction. The dataset features explicit conformational minima annotations, GCN-based atomic encodings, and carefully curated train/validation/test splits to facilitate robust model evaluation. This Zenodo record archives the complete FD25 dataset snapshot, model weights, and essential post-processing utilities. The associated GNN codebase is publicly available on GitHub. We encourage researchers to use and cite FD25 as a benchmark for molecular representation learning, transferability studies, and large-scale quantum chemistry modeling.