Produced by DeFoG's exact moses_dataset.py process + filter_dataset=True
(build_molecule_with_partial_charges -> mol2smiles, single-fragment check), symbol-only
over 8 atom types with aromatic bonds, no explicit H. Molecules whose graph fails to
rebuild to a valid single-fragment SMILES (~10.5%, all pyrrole/imidazole [nH] kekulize
failures) are dropped — so this is the DeFoG filtered survivor set. Targets
(logP, qed… See the full description on the dataset page:
https://huggingface.co/datasets/nico8771/moses_test_defog.
