HIP-OrbMol: 10.8M training and 200k validation geometries
This release uses one fixed geometry split for both UMA and OrbMol labels. The target contains 10,800,000 training and 200,000 validation samples.
Current release status
The split manifests are authoritative. The complete packed release passes exact coverage and checksum checks. All 11 million autograd labels are available. The catalogs record their exact split coverage.
The former one-million validation collection now supplies 800,000 training additions and 200,000 held-out samples.
Do not use its complete legacy validation/shards/ folder as the current validation split.
The legacy autograd shard paths are removed from this revision. Prior revisions retain them.
Fixed geometry split
| Manifest | Geometries | Parent |
|---|---|---|
manifests/sample_10m.npz |
10,000,000 | OMol-1 train, release 260123 |
manifests/training_additions_800k.npz |
800,000 | OMol-0 val-comp, release 250514 |
manifests/validation_200k.npz |
200,000 | OMol-0 val-comp, release 250514 |
manifests/train_10p8m.npz |
10,800,000 | Combined training split |
The combined manifest stores indices, natoms, and parent_partition.
Partition 0 identifies OMol-1 training rows. Partition 1 identifies OMol-0 val-comp rows.
The pair (parent_partition, indices) identifies a geometry. Parent row IDs alone are not global identifiers.
Validation keeps complete molecular-composition groups separate from training additions.
The original validation selection excludes the original training composition hashes.
Selection uses metadata coverage and domain-size quotas. It does not use model errors.
The constrained split has no inverse-probability weights.
See manifests/split_10p8m_200k.json for hashes, coverage counts, and selection details.
Packed labels
Teacher: OrbMol-v2. Hessians use autograd and float32 values.
Each packed shard stores hessian_upper, hessian_ptr, and natoms.
The upper triangle includes the diagonal. Entries use row-major numpy.triu_indices(3 * natoms) order.
The lower triangle follows by symmetry. Gzip and byte shuffle are lossless.
Energy units are eV. Force units are eV/Angstrom. Hessian units are eV/Angstrom^2.
Current packed paths:
- Original training labels:
shards/*/shard_*.h5. - Training additions:
training_additions/shards/*/shard_*.h5. - Validation:
validation_200k/shards/*/shard_*.h5.
Files under catalogs/ record published shard hashes and available row IDs.
An incomplete collection can publish immutable snapshots of its completed records.
Later snapshots can replace those files with more completed records.
Count each geometry once. Never combine a legacy shard and its packed replacement.
The original 10M training shards already use the upper-triangle format.
Provenance and other data
Geometries come from OMol25. The teacher is UMA or OrbMol-v2, as identified above. Existing DFT benchmark files remain separate from this autograd split. Existing license and attribution files continue to apply.
- Downloads last month
- 101