HIP-OrbMol-OMol25

This dataset contains OrbMol-v2 teacher predictions for 10 million OMol-1 training geometries and one million OMol-0 validation geometries. Each sample contains energy, forces, and an autograd Hessian.

Source and scope

• Source dataset: OMol25, OMol-1 training archive, release 260123.
• Validation source: OMol-0 validation val-comp, release 250514.
• Teacher model: OrbMol-v2.
• Software: orb-models==0.7.0, torch==2.8.0.
• Precision: float32-high. The dataset stores teacher arrays as float32.
• Hessian method: grouped batched force-Jacobian calculation. The calculation makes each Hessian symmetric.
• Stress labels: none.
• Atom range: 2–350 atoms.

Licenses and notices

This release has components with different terms. The dataset card uses license: other.

• OMol-1 and OMol-0 source geometries and source fields remain under CC BY 4.0. Credit Meta FAIR Chemistry and the OMol25 authors. We identify the source splits above. We selected and reformatted these records and added model labels.
• We license the generated energy, force, and Hessian labels under Apache-2.0. This is our chosen license for these generated labels. It is not a license requirement from Orbital Materials.
• OrbMol-v2 is an Orbital Materials model under Apache-2.0. A checkpoint that includes its original weights is also offered under Apache-2.0. The checkpoint notice must retain Orbital Materials' attribution and mark our fine-tuning changes. See LICENSES/Apache-2.0.txt and NOTICE.

These terms do not replace the OMol source terms. They do not grant rights to use Orbital Materials' names or marks as an endorsement.

We use the same 10 million geometries as for HIP-UMA. Its SHA-256 hash is 2587870b27c7aac29684e787f72038e9361d81fc647d7620fa8698e418b5215e.

Use ASEDatabase.get_atoms(index) with the indices from our manifests:

with np.load("manifests/sample_10m.npz") as m:
    train_atoms = omol1_train_db.get_atoms(int(m["indices"][0]))
with np.load("manifests/validation_1m.npz") as m:
    val_atoms = omol0_val_comp_db.get_atoms(int(m["indices"][0]))

Files

README.md
manifests/sample_10m.npz
manifests/sample_10m.meta.json
manifests/validation_1m.npz
manifests/validation_1m.meta.json
metadata/release.json
metadata/validation_release.json
metadata/review_checklist.md
LICENSES.md
NOTICE
LICENSES/Apache-2.0.txt
checksums.sha256
shards/000/shard_00000.h5
...
validation/shards/000/shard_00000.h5
...

HDF5 format

Each shard stores molecules without padding, with these fields:

• omol_index, natoms, atomic_numbers, coords, and atom_ptr.
• charge, spin, spin_multiplicity, and unrestricted.
• energy, forces, hessian_upper, and hessian_ptr.
• Source identity fields and validation selection fields, when present.

The spin and spin_multiplicity fields store spin multiplicity, not the spin quantum number. The dataset uses these units:

• Energy: eV.
• Coordinates: Å.
• Forces: eV/Å.
• Hessians: eV/Ų.

The Hessian uses Cartesian coordinate order (x1, y1, z1, x2, ...). The dataset stores only the upper triangle, including the diagonal, in row-major order. For N atoms, the Hessian dimension is d=3N, i.e. each sample stores d*(d+1)//2 float32 values. The stored Hessians do not use mass weighting. The labels include the teacher model's Coulomb contribution and charge response. The calculation does not subtract a fixed-charge Coulomb term.

The code below reconstructs the full Hessian:

import h5py
import numpy as np

with h5py.File("shards/000/shard_00000.h5", "r") as shard:
    sample = 0
    a, b = shard["atom_ptr"][sample:sample + 2]
    c, d = shard["hessian_ptr"][sample:sample + 2]
    n = int(shard["natoms"][sample])
    coordinates = shard["coords"][a:b]
    row, col = np.triu_indices(3 * n)
    hessian = np.empty((3 * n, 3 * n), dtype=np.float32)
    hessian[row, col] = shard["hessian_upper"][c:d]
    hessian[col, row] = hessian[row, col]