niklastr/tidy_mdqm9nc

tidy-MDQM9nc

This release contains a strictly filtered, atom-order-correct subset of MDQM9-nc for training and evaluating transferable molecular samplers.

Split	Original molecules	Retained molecules	Rejected molecules	Retained frames	Rejected frames
Train	12,306	11,224	1,082	179,584,000	17,312,000
Val	100	7	93	112,000	1,488,000
Test	100	95	5	3,420,000	180,000

Layout and schemas

Training WebDataset

Training data are in webdatasets/single_frames/train/0000.tar through 1968.tar. Each retained sample has two members named <molecule_id>_<frame_id>:

• .bin: raw little-endian float32 data shaped (N, 4) with columns (Z, x, y, z);
• .smiles: UTF-8 explicit-hydrogen, atom-mapped SMILES.

Coordinates are in nanometres. Atom map i refers to row i - 1 of the binary array. The original sample keys and shard assignment are preserved.

Validation and test trajectories

Validation and test data are one NPZ per retained molecule under trajectories/val/ and trajectories/test/. Trajectory ordering and all original arrays are preserved; only the SMILES value is replaced by the validated atom-mapped representation. The positions array has shape (frames, N, 4) and columns (Z, x, y, z) in nanometres.

Topology

topology/mdqm9-nc.sdf is the source SDF topology collection. Contrary to the previous dataset card, this repository does not contain GAFF topology tar archives.

Provenance and verification

• Source dataset: MDQM9-nc, Zenodo record 10579242.
• Source Hugging Face revision: 64727bf79c5bef7d2411598622da4531d74700f8.
• Filter implementation: strict-mdqm9nc-v1.
• Source simulations: non-cyclic QM9 molecules in vacuum near room temperature using GAFF.
• QC and output manifests: qc/.
• Per-file SHA256, byte sizes, frame counts, and shard counts: qc/release_manifest.json.

All retained training samples were copied from the atom-mapped local WebDataset only after the strict molecule-level decision. Every output shard was parsed and checked for paired members, valid tar checksums, exact mapped SMILES, expected atomic-number order, unique sample keys, and absence of rejected molecules. Every retained validation/test frame was checked for finite coordinates and exact atomic-number order before serialization.

Limitations

Graph inference from coordinates is geometry-sensitive. The strict policy intentionally prefers precision over coverage and rejects an entire molecule for any graph, mapping, or cross-shard ambiguity. This can change the chemical composition of each split. Consult the QC manifest before comparing results against the original MDQM9-nc split.

Citation

@dataset{viguera_diez_2024_10579242,
  author = {Viguera Diez, Juan and Olsson, Simon},
  title = {MDQM9-nc dataset},
  year = {2024},
  publisher = {Zenodo},
  doi = {10.26434/chemrxiv-2023-sx61w}
}