We benchmark on the community-standard Dalke NN dataset (1,000 high-similarity ChEMBL pairs) — the same dataset widely used by RDKit, CDK, and the academic MCS literature. Identical SMILES input, same ...