CrysMTM v2

A TiO₂ dataset with 63 original structures and 9,738 archived views, exposed as standard Hugging Face datasets with corrected, disjoint parent-level splits.

from datasets import load_dataset

ds = load_dataset("kurbanintelligencelab/CrysMTM", "core", revision="v2.0.2")
print(ds["train"][0]["targets"])

Install datasets[vision] for decoded images. Select augmented for all archived views. Both configurations use the same temperature-v2 protocol. Supported splits are train, validation, test_id, and test_ood; streaming is supported.

Configuration	Train	Validation	ID test	OOD test	Total
core	24	6	15	18	63
augmented	1416	354	1380	6588	9738

Contents

Each record contains IDs, phase, temperature (K), species, atomic numbers, positions (angstrom), XYZ text, an embedded image, text, nine named targets with source units, source hashes, a rotation matrix and its center. core also contains paper_rotations, the recovered matrices for replaying every archived orientation. Each parent's rotated views share its physical labels.

Electronic targets HOMO, LUMO, Eg, Ef, Et and Eta are in eV as labeled by the source. disp and bond are in Å, and vol in ų. Structures use Cartesian XYZ coordinates. text_kind=legacy-descriptors-unrevalidated identifies original archive descriptions. Source descriptor definitions are documented in the paper; v2 validation covers data integrity and rotations. Select task inputs separately from labels and fit normalization on training parents. Targets are returned in their source units.

Splits

• Train: 150, 200, 300, 400, 500, 600, 700, 850 K.
• Validation: 350, 550 K.
• ID test: 250, 450, 650, 750, 800 K.
• OOD test: 0, 50, 100, 900, 950, 1000 K.

Every temperature assignment applies to all three phases. Every view remains with its parent. The original repository's training and OOD lists overlapped at 0, 50 and 100 K; the README's 13,032 total counted overlapping memberships. This release contains 9,738 distinct view IDs. Report benchmark results with their protocol version. This release provides the temperature-v2 data and split definitions.

Custom generation

The CrysMTM toolkit supports arbitrary rotation counts per parent, seeds, explicit Euler angles or matrices, optional rendering, export and a paper orientation preset. The preset replays the recovered archive matrices around the origin, including identity, to coordinate precision. Generated images and text use the v2 renderer and geometric template. Use augmented for the original archived XYZ files, images and descriptions. Install the versioned wheel with pip, or download the source distribution. The toolkit source is available in the GitHub repository.

pip install "https://huggingface.co/datasets/kurbanintelligencelab/CrysMTM/resolve/v2.0.2/toolkit/crysmtm-2.0.2-py3-none-any.whl"
crysmtm fetch --output ./data
crysmtm generate ./data --rotations-per-structure 1000 --seed 42 --output ./generated

Generation defaults to XYZ/coordinates and geometric text. Install the optional rendering extra from the wheel to generate images. See the source README for rendering, explicit angles/matrices, phase/temperature filters and the paper preset.

Provenance and license

Converted from Figshare v1, archive MD5 74315f19453b7e2f0f76855046707e81. SHA-256 and the full validation summary are in audit.json. The data retains CC BY 4.0 and attribution to Can Polat, Erchin Serpedin, Mustafa Kurban and Hasan Kurban. The original repository snapshot is 3aa724da3c7ac64bbbe79dcf6dfbfc275fad5507. Cite the paper and pin the dataset revision used in an experiment.

Citation

@article{polat2025crysmtm,
  title={CrysMTM: a multiphase, temperature-resolved, multimodal dataset for crystalline materials},
  author={Polat, Can and Serpedin, Erchin and Kurban, Mustafa and Kurban, Hasan},
  journal={Machine Learning: Science and Technology},
  volume={6},
  number={3},
  pages={030603},
  year={2025},
  publisher={IOP Publishing}
}