Carbon-24 Unique
Please cite Martirossyan et al. (https://arxiv.org/abs/2509.12178) if your work utilizes any of the datasets referenced here.
The carbon datasets and dataset splits below have been cultivated from carbon-24 (Pickard 2020, doi: 10.24435/materialscloud:2020.0026/v1).
Totaling 4,250 unique carbon structures cultivated from the carbon-24 dataset. Duplicates were identified using StructureMatcher, which allows for improper rotations to find mappings between structures, and as such one of each enantiomorph pair was also removed in the creation of this dataset.
carbon-24_unique/carbon-24_unique_train.parquet: 2550 structures
carbon-24_unique/carbon-24_unique_val.parquet: 850 structures
carbon-24_unique/carbon-24_unique_test.parquet: 850 structures
Identical to carbon-24_unique, but split according to the number of atoms per unit cell N. The splits were done in both a low-to-high and high-to-low fashion across train, validation, and test set construction (in that order).
low_to_high/train.parquet: 2280 structures with N=6–10
low_to_high/val.parquet: 1159 structures with N=12–14
low_to_high/test.parquet: 811 structures with N=16—24
high_to_low/train.parquet: 2633 structures with N=10–24
high_to_low/val.parquet: 792 structures with N=8
high_to_low/test.parquet: 825 structures with N=6
The carbon-24_unique dataset was used with the following datasets for the 2025 NeurIPS Datasets & Benchmarks:
carbon-24_unique_with_enantiomorphs
Companion dataset to carbon-24_unique, containing enantiomorph pairs discovered within the carbon-24 dataset.
carbon_X
480 carbon structures of duplicates which have the same cell shape and same number of atoms per unit cell (N=6), with different translations X of the fractional coordinates.
carbon_NXL
353 carbon structures of duplicates which have different numbers of atoms per unit cell (N=6—16), different cell shapes L, and different translations X of the fractional coordinates.
carbon-enantiomorphs Cultivated from the carbon-24-unique-with-enantiomorphs dataset, containing: train.xyz - 80 structures which are chiral; and test.xyz - 80 of the same structures with opposite handedness.
Files in the 'supplemental' folders contain a map of material IDs from the original dataset at https://github.com/txie-93/cdvae.
perov-5_polymorph_split
A new split of the perov-5 dataset which contains 18,928 perovskite structures within which there are only 9,646 unique compositions (9,282 compositions show up twice in the dataset, while 364 compositions show up once). The new splits ensure that polymorph pairs are grouped within the training, validation, and test sets.
MP-20-polymorph-split
A new split of the MP-20 dataset which contains 45,229 structures of varying composition and structure, within which there are only 37,217 unique reduced compositions.
The new split ensures that polymorph pairs are grouped within the training, validation, or test sets, and that the N-arity of the split distributions matched that of the combined one. Materials IDs are included in the parquet files.
Alex-MP-20_Polymorph_Split A new split for performing crystal structure prediction on the Alex-MP-20 dataset (https://github.com/microsoft/mattergen/tree/main/data-release/alex-mp) which contains structures from MP-20 (Jain 2013, doi: 10.1063/1.4812323) and Alexandria (Schmidt 2022, doi: 10.24435/materialscloud:m7-50). This dataset ensures that structures of the same composition remain together in the same split.
CC-BY-4.0
Carbon-24 Unique
Please cite Martirossyan et al. (https://arxiv.org/abs/2509.12178) if your work utilizes any of the datasets referenced here.
The carbon datasets and dataset splits below have been cultivated from carbon-24 (Pickard 2020, doi: 10.24435/materialscloud:2020.0026/v1).
Totaling 4,250 unique carbon structures cultivated from the carbon-24 dataset. Duplicates were identified using StructureMatcher, which allows for improper rotations to find mappings between structures, and as such one of each enantiomorph pair was also removed in the creation of this dataset.
carbon-24_unique/carbon-24_unique_train.parquet: 2550 structures
carbon-24_unique/carbon-24_unique_val.parquet: 850 structures
carbon-24_unique/carbon-24_unique_test.parquet: 850 structures
Identical to carbon-24_unique, but split according to the number of atoms per unit cell N. The splits were done in both a low-to-high and high-to-low fashion across train, validation, and test set construction (in that order).
low_to_high/train.parquet: 2280 structures with N=6–10
low_to_high/val.parquet: 1159 structures with N=12–14
low_to_high/test.parquet: 811 structures with N=16—24
high_to_low/train.parquet: 2633 structures with N=10–24
high_to_low/val.parquet: 792 structures with N=8
high_to_low/test.parquet: 825 structures with N=6
The carbon-24_unique dataset was used with the following datasets for the 2025 NeurIPS Datasets & Benchmarks:
carbon-24_unique_with_enantiomorphs
Companion dataset to carbon-24_unique, containing enantiomorph pairs discovered within the carbon-24 dataset.
carbon_X
480 carbon structures of duplicates which have the same cell shape and same number of atoms per unit cell (N=6), with different translations X of the fractional coordinates.
carbon_NXL
353 carbon structures of duplicates which have different numbers of atoms per unit cell (N=6—16), different cell shapes L, and different translations X of the fractional coordinates.
carbon-enantiomorphs Cultivated from the carbon-24-unique-with-enantiomorphs dataset, containing: train.xyz - 80 structures which are chiral; and test.xyz - 80 of the same structures with opposite handedness.
Files in the 'supplemental' folders contain a map of material IDs from the original dataset at https://github.com/txie-93/cdvae.
perov-5_polymorph_split
A new split of the perov-5 dataset which contains 18,928 perovskite structures within which there are only 9,646 unique compositions (9,282 compositions show up twice in the dataset, while 364 compositions show up once). The new splits ensure that polymorph pairs are grouped within the training, validation, and test sets.
MP-20-polymorph-split
A new split of the MP-20 dataset which contains 45,229 structures of varying composition and structure, within which there are only 37,217 unique reduced compositions.
The new split ensures that polymorph pairs are grouped within the training, validation, or test sets, and that the N-arity of the split distributions matched that of the combined one. Materials IDs are included in the parquet files.
Alex-MP-20_Polymorph_Split A new split for performing crystal structure prediction on the Alex-MP-20 dataset (https://github.com/microsoft/mattergen/tree/main/data-release/alex-mp) which contains structures from MP-20 (Jain 2013, doi: 10.1063/1.4812323) and Alexandria (Schmidt 2022, doi: 10.24435/materialscloud:m7-50). This dataset ensures that structures of the same composition remain together in the same split.
CC-BY-4.0
