Organic-OMol1

The organic subset of the OMol-1 Train All split of Meta's Open Molecules 2025 (facebook/OMol25), with molecular identifiers added. OMol25 ships coordinates, energies and forces without SMILES, InChI or InChIKey, so its records cannot be searched or joined by compound. The records here carry those identifiers.

24,023,732 structures containing only B, Br, C, Cl, F, H, I, N, O, P and S. These are 1,090,506 molecules, 22.03 structures per molecule on average. Every structure is a single-component, carbon-containing, neutral, closed-shell singlet.

Questions and feedback

Have a question, found an issue, or used Organic-OMol1 in your research? We welcome your feedback and would love to hear about your work.

Join the discussion.

Uses

What the identifiers let you do that OMol25 alone does not:

• Group every structure of a molecule, and split by molecule so that structures of the same molecule do not land on both sides of a train and test split.
• Join DFT records to experimental measurements and public compound databases. Join on inchikey_skeleton when the experimental source has no consistent stereochemistry, which is the common case. Join on inchikey_stereo when stereochemistry matters, at lower coverage.
• Do not join on formula. Constitutional isomers share a formula. Formula matching was tested on this corpus and fails.
• Train interatomic potentials and 3D property models on coordinates, energies and forces evaluated at one level of theory.
• Featurize the same records as SMILES or molecular graphs for models that do not take coordinates, including sequence and graph architectures.
• Select configurations by force before training, using forces_flat.

A Python package for compound matching and molecular feature generation against this release is planned as a separate public repository.

Chemical identifiers added

column	what it holds
canonical_smiles	canonical isomeric SMILES
smiles_flat	SMILES without stereochemistry
smiles_stereo	SMILES with stereochemistry
inchi	Standard InChI
inchikey_stereo	full Standard InChIKey. Encodes isotope, charge and protonation state. Does not encode spin
inchikey_skeleton	first InChIKey block, the connectivity skeleton. Collapses stereoisomers, and can collapse mobile-hydrogen tautomers
inchikey_stereo_geometric	full Standard InChIKey with stereochemistry taken from the coordinates instead of the graph
formula	molecular formula

Meta Data

111 in all, counted from the released Parquet schema. 82 were produced by this work and do not exist in OMol25. The other 29 carry OMol25's values unchanged. No new quantum chemistry was run. Every energy, force, orbital energy and partial charge is OMol25's calculation. A *_status column over an OMol25 property counts as produced here, because it records whether their value is present rather than holding a value of its own.

group	columns	added here	from OMol25
Availability flags	23	23	0
Identity evidence	15	15	0
Composition and scope	13	13	0
DFT scalars	13	0	13
Per-atom arrays	12	0	12
Parsed inputs with validity	9	9	0
Record provenance	9	5	4
Chemical identifiers	8	8	0
Force and sampling strata	6	6	0
Units and geometry bookkeeping	3	3	0

Each physical property has a *_status column saying whether that property is present and usable for that structure. Filter on it before regressing on a value. A null is not a zero.

How to load it

One table, structures/. 200 Parquet files, 24,023,732 rows, 111 columns. Each row is one structure with its coordinates, energies, forces and DFT properties, and the identifier columns on the same row.

from datasets import load_dataset

ds = load_dataset("ahmed-lab-umich/Organic-OMol1", streaming=True)["train"]

To search or join by structure without moving the geometry, read only the columns you need. Parquet is columnar, so this touches a small fraction of the 48 GiB:

import pyarrow.parquet as pq
ids = pq.read_table(
    "hf://datasets/ahmed-lab-umich/Organic-OMol1/structures/",
    columns=["inchikey_skeleton", "inchikey_stereo", "canonical_smiles",
             "formula", "evidence_tier"])

Finding every structure of a molecule

Filter on an identifier column. inchikey_skeleton gives the connectivity match, which is what joins to most experimental sources. inchikey_stereo together with charge and spin gives the exact species.

same = ds.filter(lambda r: r["inchikey_skeleton"] == "ABCDEFGHIJKLMN")

What to do with the structures of one molecule is yours to decide: take them all, take the lowest-energy one, or filter on force first.

Coordinates and forces are flat. positions_flat is a vector of length 3N in row-major order, atomic_numbers a parallel vector of length N.

import numpy as np
xyz = np.asarray(row["positions_flat"], dtype=float).reshape(-1, 3)  # angstrom
Z   = np.asarray(row["atomic_numbers"], dtype=int)

Provenance

The source records were obtained from the OMol25 electronic-structure release, distributed through Meta's OMol25 repository and hosted by Argonne National Laboratory through Globus. No OMol25 model checkpoints are included.

Every structure carries raw_uid, data_id and source, resolving to its position in the OMol25 archive. source is OMol25's own identifier string, carried through unchanged.

The underlying calculations are OMol25's, at the ωB97M-V/def2-TZVPD level in ORCA 6.0.0. This release covers only the OMol-1 Train All split and redistributes no OMol25 validation or test records. Check identity overlap yourself before evaluating against an OMol25 benchmark.

Acknowledgment

This research was supported by the Electric Vehicle Center, University of Michigan, through funding provided by the State of Michigan.

Citation

@misc{organic_omol1,
  title     = {Organic-OMol1: Reconstructed Molecular Identities for the
               Organic Subset of Open Molecules 2025},
  author    = {Ahmed, Alauddin},
  year      = {2026},
  publisher = {University of Michigan},
  version   = {1.0.0},
  url       = {https://huggingface.co/datasets/ahmed-lab-umich/Organic-OMol1}
}

@misc{levine2025openmolecules2025omol25,
  title         = {The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models},
  author        = {Daniel S. Levine and Muhammed Shuaibi and Evan Walter Clark
                   Spotte-Smith and Michael G. Taylor and Muhammad R. Hasyim and
                   Kyle Michel and Ilyes Batatia and G\'abor Cs\'anyi and Misko
                   Dzamba and Peter Eastman and Nathan C. Frey and Xiang Fu and
                   Vahe Gharakhanyan and Aditi S. Krishnapriyan and Joshua A.
                   Rackers and Sanjeev Raja and Ammar Rizvi and Andrew S. Rosen
                   and Zachary Ulissi and Santiago Vargas and C. Lawrence Zitnick
                   and Samuel M. Blau and Brandon M. Wood},
  year          = {2025},
  eprint        = {2505.08762},
  archivePrefix = {arXiv},
  primaryClass  = {physics.chem-ph}
}

@article{orca6,
  title   = {Software update: The ORCA program system -- Version 6.0},
  author  = {Neese, Frank},
  journal = {WIREs Computational Molecular Science},
  volume  = {15},
  number  = {2},
  pages   = {e70019},
  year    = {2025},
  doi     = {10.1002/wcms.70019}
}

License

Organic-OMol1 is released under CC BY 4.0. The OMol25 records it derives from are also CC BY 4.0. Meta's FAIR Chemistry License applies to the OMol25 model checkpoints, which are not included here.