neurips-dataset/QHMat_sample_tar

QHMat Dataset Sample (1 GB) Preview

Dataset Details

It contains a ~1 GB subset derived from the full QHMat dataset, intended for quick data quality checks. The complete public release of full QHMat (~4.1TB) is planned after peer review. This subset shares the same indexing conventions and data format as the full release.

Hierarchy: Full QHMat (4.1 TB) → 1 GB subset (this repository)

Dataset Description

This dataset is a small, byte-limited preview intended for peer review, sanity checks, and pipeline testing. It contains 45 crystal/material structures as serialized graph/tensor records (Hamiltonian and overlap blocks, geometry, and metadata).

Files in this release

File / pattern	Description
train-000000.tar, train-000001.tar, …	WebDataset-style archives. Each logical sample has two members sharing the same basename.
00000000.pkl, …	Pickle blob of torch_geometric.data.Data (same bytes as stored in the source LMDB value).
00000000.json, …	Small JSON sidecar: lmdb_key (integer database index), value_bytes (byte length of .pkl).
train-manifest.csv	Table mapping sample_index → shard file, member names, lmdb_key, value_bytes.

Statistics (this build)

• Samples: 45 (rows in train-manifest.csv).
• Shards: 2 (train-000000.tar, train-000001.tar).
• Approximate on-disk tar footprint: ~568 MiB + ~490 MiB (see files after upload; exact sizes depend on filesystem).

Example record fields

After pickle.loads on a .pkl file, typical Data attributes include (names may vary slightly by pipeline version):

• Geometry / graph: pos, atoms, lattice, multi_edge_index, multi_edge_vec, lattice_translation_vector, …
• Hamiltonian / overlap: diagonal_hamiltonian, off_diagonal_hamiltonian, optional diagonal_overlap, off_diagonal_overlap, with associated *_mask tensors.
• Metadata: database_idx, mp_id, raw_basename, fermi_level, elapsed_time, …

Use torch_geometric introspection or print data.keys after loading.

Minimal load example (single sample)

import io
import json
import pickle
import tarfile

import torch  # noqa: F401 — needed for unpickling tensors inside Data

tar_path = "train-000000.tar"
with tarfile.open(tar_path, "r") as tar:
    pkl_info = tar.getmember("00000000.pkl")
    pkl_bytes = tar.extractfile(pkl_info).read()
    meta = json.loads(tar.extractfile(tar.getmember("00000000.json")).read())

data = pickle.loads(pkl_bytes)
print(meta)   # {'lmdb_key': ..., 'value_bytes': ...}
print(data)   # torch_geometric.data.Data

Direct Use

• Training or benchmarking graph neural networks and related models on Hamiltonian / overlap tensor blocks paired with crystal structure fields.
• Dataset loader development and integration tests before downloading larger subsets.
• Reproducibility checks during peer review (inspect manifests, decode samples).

Out-of-Scope Use

• Not a guaranteed statistically representative sample of the full corpus (structure count is small).
• Not for safety-critical or high-stakes decisions without domain validation.
• Not a substitute for the full release when training production-scale models.

Splits

• Single split: all samples are shipped under the train-* shard prefix (preview bundle). There is no separate val/test split in this artifact.

Data Collection and Processing

1. Upstream: Structures and electronic-structure-derived tensors originate from computational workflows (OpenMX-class outputs) merged into a large LMDB in the parent project.
2. Subset: A ~1 GiB LMDB subset was built by copying records in canonical index order until the byte budget was reached (same convention as documented in the upstream repo).

Personal and Sensitive Information

This dataset is not expected to contain personal data. Identifiers such as mp_id refer to materials database IDs, not individuals.

Bias, Risks, and Limitations

• Coverage: The preview contains very few structures relative to the full corpus; metrics are not indicative of full-dataset diversity.
• Simulation bias: Electronic-structure workflows, basis choices, and convergence criteria induce systematic differences versus experiment.
• Technical: .pkl relies on Python pickling; loading requires compatible torch / torch_geometric versions with the training stack used to build the data.

Recommendations

• Validate loader behavior on this preview before scaling to multi-TB subsets.
• Pin dependency versions in your training repo.
• Prefer citing the paper + dataset version + manifest checksum once available.