equiv_dens_ml datasets

Training labels and geometries for the equiv_dens_ml equivariant density-prediction project (paper-era DenSNet models plus the JACS revision campaign: water clusters, malonaldehyde, QM7-X, and OMol25/CSH permutation work).

What is here

Path	Content	Approx. size
*.npy (root)	Paper-era labels: energies, orbital/DF basis tables, small-molecule (H2O, ethanethiol, resorcinol, thiophene n-mers, QM7-X) DF coefficients	~11 GB
revision/water_labels/, revision/water_clusters*/, revision/water_ood*/	Water-cluster (n=2-24) geometries and PBE-D4 / ωB97M-V labels, including MD-expansion and new-motif waves	~7.6 GB
revision/malonaldehyde*/, revision/malonaldehyde_labels/	Proton-transfer scan geometries and labels	~360 MB
revision/csh_npy/, revision/csh_df/	OMol25 CSH shard conversions (Fock -> DF coefficients) used to validate the AO-permutation pipeline	~320 MB
revision/omol_csh/manifests/, revision/omol_csh/ao_permutation*.json	Authored frame manifests and the solved AO-permutation map for OMol25 CSH; not the raw OMol25 electronic-structure HDF5 (see Excluded below)	~1.6 MB
revision/qm7x/raw/*.hdf5	Raw QM7-X single-point HDF5 shards used as ORCA campaign input	~77 GB
revision/qm7x/qm7x_official_8000.jsonl	Official QM7-X 8000-molecule index	small
revision/ethanol_dft_shards/, revision/water_102/, revision/water_pilot/	Intermediate campaign label sets	~580 MB
revision/acac/, revision/crambin_frags/, revision/meoh_water/, revision/nma_dimer/, revision/paracetamol/, revision/pgg/	H-bond new-systems geometry seeds	small
revision/ood/, revision/pbe0/	Ethanol OOD set and PBE0 subset	~30 MB
revision/sad_*.npy	Free-atom superposition-of-atomic-densities (SAD) priors per theory level	small

Total hosted here: ~103 GB.

Excluded (not hosted on this repo)

• revision/omol_csh/*.h5 (~318 GB) - the raw OMol25 Common-Structure-Hierarchy electronic-structure dump (Fock/density matrices). Re-download via Globus/MDF per facebook/OMol25 (train_4M split); the largest shard (omol_csh_58k_train.h5, ~258 GB) exceeds the Hub's per-file limit on its own.
• revision/omol25/ (~20 GB) - the OMol25 train_4M index mirror. Fetch directly from facebook/OMol25 instead of via this repo.
• revision/*/shards*/, revision/qm7x/shards/ - per-frame ORCA job-tracking JSON manifests. Regenerated by scripts/revision/lib/build_shards_from_xyz.py; not scientific data, and one alone (qm7x/shards/train/) had >20k small files.
• Model checkpoints, training run directories, raw ORCA/MD outputs - out of scope for a dataset repo. See the project's results/revision/ for compact JSON/JSONL/MD summaries (train/test overlap, timing, calibration) checked directly into git.

Formats

• .npy: numpy.load(..., allow_pickle=True).item() dictionaries (or 2D arrays for basis tables), keyed by atom_numbers, positions, energies, forces, DF coefficients, etc. See src/equiv_dens/data/ in the code repo for the exact schema per file family.
• .xyz: standard XYZ geometry files (already duplicated in the code repo under datasets/revision/**/*.xyz for clone-without-download workflows).
• .json / .jsonl: manifests, AO-permutation maps, and the QM7-X official index.
• .hdf5: raw QM7-X single-point calculations (ASE/QM7-X native layout).

Download

pip install -U huggingface_hub
hf download muhammadhasyim/equiv_dens_ml --repo-type dataset --local-dir /path/to/equiv_dens_data

Or selectively, e.g. only the water campaign:

hf download muhammadhasyim/equiv_dens_ml --repo-type dataset \
  --include "revision/water_*" --local-dir /path/to/equiv_dens_data

See scripts/download_hf_datasets.sh in the code repository for a wrapper that also wires the download into the project's expected datasets/ layout via symlinks.

License

MIT, matching the code repository. OMol25-derived content that is excluded here retains its own upstream license/access terms; consult facebook/OMol25 before redistributing it.