Each row is a molecule as canonical SMILES plus RDKit-recomputed targets. DeFoG-faithful prep: molecules are featurized over 9 neutral atom types (charges not modeled), and each SMILES is the form its neutral graph decodes to via the partial-charge build — cations (N+) recovered, anions ([O-]) as the neutral acid. Targets are computed on this representable form so they match the graph the model sees.

Source: torch_molecule ZINC-250k… See the full description on the dataset page: https://huggingface.co/datasets/nico8771/zinc_neutral.