HIP-UMA-OMol25

UMA-S-1.2 (uma-s-1p2) autograd Hessian predictions on a subset of OMol-1 train geometries, using the omol task.

Current snapshot (2026-10-05): 24,418 complete shards, 480.8 GB, 8,818,645 samples. The goal is around 10M samples (41911 shards planned).

Shard files are grouped to stay under Hugging Face's 10,000-files-per-directory limit:

shards/00/shard_00000.h5 … shards/41/shard_41910.h5

Contents

Each HDF5 file contains molecular configurations with:

• atomic_numbers, coords, and natoms
• charge, spin, spin_multiplicity, and unrestricted
• energy, forces, and dense hessian_flat
• atom_ptr and hessian_ptr for reconstructing variable-size arrays
• OMol matching fields: omol_index, source, data_id, sid, reference_source, and composition
• optional OMol reference labels: dft_energy, dft_forces, and has_dft_labels

The teacher labels are float32:

• energy: eV
• forces: eV/Angstrom
• Hessian: eV/Angstrom^2

The Hessian is dense, symmetrized, and has shape (3N, 3N).

Downloading and selecting OMol-1 geometries

The samples are selected from the OMol-1 train release 260123. See the official OMol25 page. Download the complete 592 GB ASE-LMDB archive:

wget -c \
  https://dl.fbaipublicfiles.com/opencatalystproject/data/omol/260123/train.tar.gz
tar -xf train.tar.gz

The 10M indices that we are using are saved in the manifest. Download:

uvx --from huggingface_hub hf download \
  andreasburger/HIP-UMA-OMol25 \
  manifests/sample_10m.npz \
  manifests/sample_10m.meta.json \
  --repo-type dataset \
  --local-dir .

The indices from the manifest are designed to be used with the extracted ASE-LMDB records. Use a contiguous index range to process only one part:

import numpy as np
from fairchem.core.datasets import AseDBDataset

manifest = np.load("manifests/sample_10m.npz")
selected = manifest["indices"]
dataset = AseDBDataset({"src": "/path/to/extracted/train"})

start, stop = 0, 100_000
for omol_index in selected[start:stop]:
    atoms = dataset.get_atoms(int(omol_index))
    # Process atoms here.

Reproducing UMA inputs

Use task_name=omol, UMA-S-1.2, FP32, and a 120 Angstrom molecular cell. The OMol25 spin field is the spin multiplicity. Charge and spin are required inputs and must be passed through the equivalent of r_data_keys=["spin", "charge"].

Provenance

• Parent dataset: OMol-1 train, release 260123
• Sampling: fixed-seed uniform sample, seed 20260729
• Model: uma-s-1p2
• Hessian mode: loop and batched autograd; see each shard attribute
• Precision: float32

omol_index is the index into the parent OMol-1 train ASE-LMDB.

Format

We use HDF5. It preserves random access and variable-size dense Hessians without padding. For training, FairChem graph/training caches should be generated locally from these files because their representation depends on FairChem version and training configuration.

The verified October 5 snapshot contains only complete shards. Older partial checkpoints remain in the repository; exclude files ending in .partial.h5.