Generative Chemistry Benchmark · Insilico Medicine
MOSES
The standard benchmark for molecular generation: fixed splits, baseline models and distribution metrics.
What MOSES is
MOSES is a benchmark, not a corpus, and that distinction is the point: two generative chemistry models are comparable only if they were trained on the same molecules and scored by the same measures. It supplies a set of 1,936,962 drug-like molecules refined from a commercial screening library — filtered by molecular weight, rotatable bonds, logP, element composition and medicinal-chemistry and PAINS filters — split into roughly 1.6M training, 176k test and 176k scaffold-test molecules. The scaffold split is the interesting one: its Bemis–Murcko scaffolds appear in neither of the other sets, so it measures whether a model invents chemistry it has not seen. Alongside the data come the metrics everyone quotes — validity, uniqueness, novelty, internal diversity, fragment and scaffold similarity, nearest-neighbour similarity and Fréchet ChemNet distance — implemented once, plus trained baselines to compare against.
Get the data
Its authors publish MOSES outside the Hugging Face Hub, so the dataset explorer has no table to read and there are no rows on this page. The files are linked below, and the research agent can fetch them onto its machine and work through them there.
A copy you can read row by row
A Hub copy of the full molecule table (1,936,962 rows) as SMILES and InChI, browsable row by row. It is a single split under CC BY 4.0 — not the train/test/scaffold split a reported benchmark number requires.
Try asking
“Download the MOSES test and scaffold-test splits onto the machine, compute each molecule's Bemis–Murcko scaffold with RDKit, and check that no scaffold appears in both.”
Licence
Commercial use permitted · MIT
Read 2026-10-24: the repository that publishes the splits, the metric implementations and the baselines declares MIT. The Hub mirror of the molecule table is published separately under CC BY 4.0, which requires attribution.
Questions
Can I browse MOSES in the browser?
Not in full. Its authors publish it outside the Hugging Face Hub, so the dataset explorer has no table to read (format: CSV splits with precomputed reference statistics, and a pip-installable metrics package). A copy is on Hugging Face and its rows open in the explorer; what the copy leaves out is set out under “Get the data”.
What licence is MOSES under?
MIT. Read 2026-10-24: the repository that publishes the splits, the metric implementations and the baselines declares MIT. The Hub mirror of the molecule table is published separately under CC BY 4.0, which requires attribution.
How to cite it
Polykovskiy, D. et al. Molecular Sets (MOSES): a benchmarking platform for molecular generation models. Frontiers in Pharmacology 11, 565644 (2020).