GPSK-Inorganic-Crystals

A flat, joint-balanced corpus of 2,000,115 inorganic crystal structures with conditioning metadata (space group, crystal system, magnetic ordering, formation energy, band gap), drawn from public DFT and experimental databases and curated to remove the heavy class skew of the raw merge.

At a glance

Rows	2,000,115
Format	Apache Parquet, 41 chunks × 50,000 rows
Schema	21 columns (full CIF + metadata)
Size on disk	1.1 GB
Coverage	230 space groups, 7 crystal systems, 89 chemical elements
Cell range	1 – 500 atoms / unit cell (mean = 14.7, p99 = 96)
License	CC-BY-4.0 (union of source licenses)

Quick start

from datasets import load_dataset

ds = load_dataset("willgbryan13/gpsk-inorganic-crystals", split="train")
print(ds)
print(ds[0]["formula"], ds[0]["space_group"], ds[0]["formation_energy"])

Or directly with pandas:

import pandas as pd
from pathlib import Path

df = pd.concat(
    [pd.read_parquet(p) for p in sorted(Path("chunks").glob("chunk_*.parquet"))],
    ignore_index=True,
)
print(df.shape)         # (2000115, 21)

What's in it

Every row is one inorganic crystal structure with a full CIF and a fixed set of conditioning fields. The corpus is balanced across (space group × n_sites bin × chemical family) so that the long tail of rare families is not crushed by the dominant ones: every entry from rock_salt_oxide, trigonal_AFM_oxide, olivine_phosphate, perovskite_distorted, L1₀_tetragonal, hex_RE_TM, rock_salt_halide, and hex_binary_chalcogenide (the eight rarest families) is preserved exactly. The dominant families are uniformly subsampled to match.

Cell-skew ratio: 3,084 × (vs 583,109 × in the raw merge).

Sources

Source	License	Provenance	Notes
Alexandria PBE	CC-BY-4.0	DFT-relaxed (PBE)	bulk
OQMD	CC-BY-4.0	DFT-relaxed (PBE)	bulk + targeted slices
Materials Project	CC-BY-4.0	DFT-relaxed (PBE)	bulk + AFM/FiM + large-cell targeted slices
COD	CC-BY-4.0	Experimental (XRD)	inorganic-only
MAGNDATA (Bilbao)	open	Experimental (magnetic)	experimentally-verified AFM / FM / FiM labels

All structures are released under CC-BY-4.0 (the most-restrictive license shared across all included sources). The per-entry license column carries the originating tag so downstream users can filter by license requirements. No paywalled databases (no ICSD) are redistributed.

Schema

Each parquet row has the following columns:

Column	Type	Description
material_id	str	Source-prefixed unique ID
fingerprint	str	Entalpic BAWL fingerprint where available
cif	str	Full CIF text
formula	str	Unit-cell composition
reduced_formula	str	Reduced formula
n_sites	int	Atoms in unit cell
nelements	int	Distinct chemical elements
crystal_system	str	spglib-derived: cubic / tetragonal / trigonal / hexagonal / orthorhombic / monoclinic / triclinic
space_group	int	spglib-derived, 1–230
band_gap	float	DFT band gap, eV (NaN where unknown)
formation_energy	float	Per-atom formation energy, eV (NaN where unknown)
e_above_hull	float	Hull distance, eV/atom (0 where unknown)
magnetic_ordering	str	NM / FM / AFM / FiM
magnetization	float	Total magnetization, μB
family	str	Heuristic chemical-family label (see below)
source	str	Originating data source
license	str	Per-entry license tag
provenance	str	dft_relaxed / experimental / dft_initial
functional	str	PBE / PBEsol / SCAN / experimental
original_id	str	Source-native ID

Family taxonomy

The family column is a coarse rule-based classifier intended as a stratification handle, not a definitive crystallographic taxonomy:

Group	Family labels
Rock-salt	rock_salt_oxide, rock_salt_halide
Perovskite	perovskite_cubic, perovskite_distorted
L1₀	L10_tetragonal
Hexagonal	hex_RE_TM, hex_binary_chalcogenide
Trigonal	trigonal_AFM_oxide, trigonal_binary
Spinel / complex oxide	spinel_or_complex_oxide
Olivine	olivine_phosphate
Intermetallic	intermetallic_binary, intermetallic_complex
Other oxides	non_TM_oxide, TM_oxide_other
Halides	halide_other
Catch-all	other

Quality control

• Coarse dedup by (reduced_formula, space_group, n_sites_bin) hash
• Precise dedup by pymatgen StructureMatcher (ltol=0.2, stol=0.3, angle_tol=5°) on coarse-hash collisions
• Disordered / partial-occupancy structures rejected at parse time
• Cell size restricted to [1, 500] atoms
• Symmetry assigned with spglib (symprec=0.1)
• Magnetic ordering derived from per-site moments for DFT sources; carried through verbatim from MAGNDATA for experimental magnetic entries

Distribution snapshots

Magnetic ordering	NM 1,586,882 / FM 379,861 / FiM 24,053 / AFM 9,319
Crystal system	orthorhombic 27% / tetragonal 19% / monoclinic 18% / trigonal 15% / cubic 9% / hexagonal 7% / triclinic 5%
n_sites (mean / p95 / p99 / max)	14.7 / 30 / 96 / 500
Cell volume (p50 / p95 / p99)	239 ų / 1,005 ų / 2,093 ų
e_above_hull (p50 / p95)	0.000 / 0.020 eV/atom

Reproducibility

The corpus is deterministic given fixed source snapshots: re-running the pipeline (with the same input snapshots and the seeds recorded in manifest.json) reproduces it byte-for-byte. The curation procedure is described in manifest.json along with all binning parameters and the per-cell cap.

Considerations for using the data

Biases. The corpus inherits the biases of its underlying sources:

• DFT-relaxed structures are computed with PBE (and minor PBEsol/SCAN portions); GGA-PBE is known to under-predict band gaps and over-stabilize delocalized states. The band_gap column should be treated as a relative ordering, not absolute.
• Magnetic-ordering labels for DFT sources are derived from per-site moments and may misclassify weakly magnetic or noncollinear materials. Experimental labels from MAGNDATA (carried with source = magndata) are the highest-confidence subset.
• Curation balancing is on (space_group, n_sites bin, family) only; chemistry-level imbalance (e.g., oxide-heavy composition) is not corrected.
• AFM coverage is 9,319 entries (0.47% of corpus) — sparse relative to NM/FM. Models trained for AFM conditioning will likely benefit from upweighting this band.
• Cell-size cap of 500 atoms excludes large MOFs, complex Heusler supercells, and many interface structures. Mean cell size is 14.7 atoms; tails are real but thin.

Out-of-scope. This corpus contains only inorganic crystals. It is not suitable for: organic molecular crystals, 2D materials (no C2DB), amorphous structures, surfaces or interfaces, or anything with partial occupancy.

Citation

If you use this dataset, please cite both this corpus and the originating source databases (Alexandria, OQMD, Materials Project, COD, MAGNDATA). A CITATION.cff file is provided alongside this card.

License

This corpus is released under CC-BY-4.0. Individual entries carry the per-source license in the license column; the union license is the most restrictive across all sources.