liuganghuggingface/moldata-qm9

QM9 equilibrium quantum molecular properties

Quantum-chemical properties of 133,885 small organic molecules, with molecular structures, calculated electronic and thermodynamic targets, and corresponding 3D geometries.

Part of the Molecular Datasets collection.

Get started

from datasets import load_dataset
dataset = load_dataset("liuganghuggingface/moldata-qm9", "default")

Where the data comes from

• Original data: https://doi.org/10.6084/m9.figshare.c.978904.v5
• Original study: Ramakrishnan et al. Quantum chemistry structures and properties of 134 kilo molecules (2014)
• Redistributed by: MoleculeNet (DeepChem) (dataset page)

Tables

Config	Splits	Rows	Targets
default	full	133,885	A Rotational constant A (regression; GHz), B Rotational constant B (regression; GHz), C Rotational constant C (regression; GHz), mu Dipole moment (regression; Debye), alpha Isotropic polarizability (regression; Bohr^3), homo HOMO energy (regression; Hartree), lumo LUMO energy (regression; Hartree), gap HOMO–LUMO gap (regression; Hartree) and 11 more

data/ contains rows with readable molecular structures; identical observations are collapsed. observations/ contains all source rows.

Source label columns are text, including values such as <10. SMILES_original retains the input representation; SMILES_canonical provides a canonical representation. structure_key groups molecular components, and structure_status records parsing outcomes.

Files and modalities

Modality	Where	Content
3D structures	structures/qm9.tar.gz (53 MB)	DeepChem QM9 archive, unchanged: qm9.sdf holds the B3LYP/6-31G(2df,p) equilibrium geometry of each molecule, and qm9.sdf.csv its properties. Take the first tab-separated field of the SDF record title (gdb 1) and replace its space with an underscore to match mol_id (gdb_1).

Download an individual file:

from huggingface_hub import hf_hub_download
path = hf_hub_download("liuganghuggingface/moldata-qm9", "structures/qm9.tar.gz", repo_type="dataset")

Data quality

• computed: Labels are computed (for example DFT or molecular dynamics).
• all structures readable: All 133,885 SMILES were read with RDKit.
• 72% undefined stereo: 72% of molecules have a stereocenter whose configuration is not given, so mirror-image forms cannot be told apart.
• R² limit 0.99: Across 1,653 cases of a molecule measured more than once, the spread between repeats is under 0.01 times the spread of all values, so even a perfect model reaches an R² of about 0.99. Part of the spread can come from conditions the source does not record.
• 100% measured per task: In a typical task, 100% of molecules have a value. A missing value means not measured, not inactive.
• 3,004 extreme values: 3,004 of 2,543,815 values lie more than 6 robust standard deviations from the median of their target (on a log scale when values span 1,000-fold or more). They are often typing or unit mistakes, and sometimes real; look at them before training.
• 133,798 molecules per task: A typical task has 133,798 labeled molecules.
• no source split: Make your own split. A random split puts near-identical molecules in both training and test sets, so results look better than they will on new compounds. Splitting by scaffold, or by date when dates are given, gives a more realistic estimate.

Dataset notes

• QM9 contains a documented uncharacterized-structures list. DeepChem fixed incorrect formal charges in older SDF snapshots in 2026; CSV and SDF should be aligned by molecule ID. Units differ across tasks; atomization and total energies are distinct.
• All 133,885 records are included. The source_uncharacterized_geometry column marks the 3,054 molecules whose geometry the source flags as inconsistent.

License

CC0-1.0

Citation

Cite the original study and data source:

• Ramakrishnan et al. Quantum chemistry structures and properties of 134 kilo molecules (2014)
• Wu et al. MoleculeNet: a benchmark for molecular machine learning (2018)
• Official DeepChem loader and task definitions
• Original QM9 data record and explicit CC0 data license
• Original list of 3054 QM9 structures failing geometry consistency checks