MOSES (MOlecular SEtS) is a benchmark for molecular generation models, combining a curated dataset, fixed data splits, standardized evaluation metrics, implementations, and baseline results. section 10.2.4
Its standardized collection contains:
- 1.9 million drug-like molecules from the ZINC Clean Leads subset. section 10.2.4
- Canonical SMILES, removed stereochemistry, stripped salts, and filtering for drug-likeness, synthetic accessibility, and molecular size. section 10.2.4
- Predefined train, test, and scaffold splits supporting fair model comparison. section 10.2.4
- Evaluation of reconstruction quality, latent-space organization, chemical validity, and generation diversity. section 10.2.4
