MV-CrystalBench

GitHub repository and evaluation results · Hugging Face dataset

MV-CrystalBench is an English, four-option visual question-answering benchmark for spatial reasoning about idealized crystal structures. It contains 716 questions over 48 base structures, plus a 200-question mini subset for lower-cost evaluation. Answers are computed and checked from atomic geometry and crystallographic symmetry, rather than generated by a language model. The release includes 1,507 PNGs referenced by these questions, each 512 × 512 pixels.

The two configurations are full (default) and mini, each with a test split. Mini is an overlapping, exact subset of full, not an independent train/test partition. There is no training split. Structures recur across tasks; do not treat questions as independent samples or claim unseen-structure generalization.

Tasks and composition

Evaluation stratum	Full	Mini	Programmatic answer source
Q1a: displayed atom count	100	25	Count displayed atoms, including periodic boundary images.
Q1b: cell stoichiometry	100	25	Count atoms assigned to the original cell after boundary sharing.
Q2 easy: local coordination	60	25	Complete isolated center-and-neighbor environment.
Q2 hard: embedded coordination	60	25	Complete neighbor shell within a cell or supercell.
Q3: relative 3D distance	100	25	Compare displayed Euclidean distances A–B and A–C.
Q4: crystal system	96	25	spglib symmetry analysis.
Q5: structure identity	100	25	Known positive pairs and controlled structural changes.
Q6: viewing direction	100	25	Strict directions with symmetry-inequivalent answer options.
Total	716	200	Four options, exactly one answer.

The catalogue covers six crystal systems: 12 cubic, 8 hexagonal, 8 monoclinic, 8 orthorhombic, 6 tetragonal, and 6 trigonal structures. It has no triclinic structures. Full-set Q4 contains 16 questions per represented crystal system. The generation and sampling seed is 20261021. Full-set correct letters are exactly balanced at 179 each; the mini set is approximately balanced. Each base structure contributes at most four questions per question type, with at most two per Q2 difficulty.

Files and image resolution

README.md
LICENSE
manifest.json
questions.jsonl
questions_mini.jsonl
images/scenes/<anonymous_scene_id>/<view_id>.png
provenance/aflow/...

The JSONL files are byte-identical to the frozen source question files. Each image_paths entry, for example scenes/s.../v00.png, is relative to the images/ directory, not the dataset root. Resolve it as dataset_root / "images" / image_path. Paths remain strings when loaded with Hugging Face Datasets; this release does not declare them as an automatically decoded Image feature. manifest.json lists SHA-256 and byte size for every other package file; it intentionally omits its own hash.

Only question-referenced PNGs are packaged. The release does not include all eight candidate views per scene, scene manifests, or per-atom visibility JSONs. It supports evaluation of the released questions, but full view ablation requires rebuilding the complete scenes and visibility metadata with the companion generation code. Extra views must not be inferred from these filenames.

Loading with JSON and Pillow

For a local checkout or downloaded snapshot:

import json
from pathlib import Path
from PIL import Image

root = Path("hf_dataset")
with (root / "questions_mini.jsonl").open(encoding="utf-8") as handle:
    questions = [json.loads(line) for line in handle if line.strip()]

question = questions[0]
images = []
for relative in question["image_paths"]:
    with Image.open(root / "images" / relative) as image:
        images.append(image.convert("RGB"))

# Only these fields, and the selected images, belong in the image prompt.
prompt_fields = {"question": question["question"], "options": question["options"]}

Loading with Hugging Face Datasets

Install the optional reader separately with pip install datasets pillow. The following local loader selects the same JSON files as the card's two configs:

from pathlib import Path
from datasets import load_dataset
from PIL import Image

root = Path("hf_dataset")
config_files = {"full": "questions.jsonl", "mini": "questions_mini.jsonl"}
config = "mini"  # Use "full" for all 716 questions.
dataset = load_dataset(
    "json",
    data_files={"test": str(root / config_files[config])},
    split="test",
)
row = dataset[0]
with Image.open(root / "images" / row["image_paths"][0]) as image:
    first_image = image.convert("RGB")

The published repository is chessnash01/mv-crystalbench. Its full and mini configs can be selected by name with load_dataset("chessnash01/mv-crystalbench", name="mini", split="test"). No dataset DOI has been assigned. Download the image files as well as the JSONL; the dataset loader does not automatically fetch and decode images referenced by these relative strings. Hugging Face's loading pipeline may insert nulls for optional metadata fields and round metadata floating-point values. Use Python's JSON reader on the original JSONL files when exact source-object or numeric preservation matters.

Fields and prompt boundaries

Field	Meaning
id	Stable question identifier.
qtype, difficulty	Task and difficulty for evaluation grouping.
structure	Internal structure identifier; never send to a model.
scene_id, view_ids	Anonymous scene/view references; Q5 can reference two scenes.
image_paths	Ordered relative PNG paths, resolved under images/.
question, options	English question and four choices keyed A–D.
answer, answer_text	Ground-truth answer letter/text; never send to a model.
meta	Internal verification and sampling details, including base_structure; never send to a model.
text_repr	Coordinate representation for the explicitly selected text condition.

Do not serialize the entire record into a prompt. In the image condition, send only the question, options, and selected images. In the text condition, send the question, options, and text_repr; do not send images. In the no_image condition, send only the question and options. Internal metadata can contain symmetry, counts, distances, or mutation labels that disclose answers. Structure/mineral names and space-group information must not be added to the prompt. Crystal-system names in Q4's answer options are intentional.

text_repr contains lattice parameters/vectors and all displayed atom elements and unwrapped fractional coordinates, with [A], [B], and [C] highlights. It excludes structure names and symmetry labels. Q1a and Q6 are treated as pure visual tasks: their text_repr is empty and meta.text_applicable is false. Thus mini image/no-image evaluation has 200 questions, while applicable text evaluation has 150. Image-minus-text comparisons must use the same 150 paired questions, not subtract accuracies from different denominators.

Generation conventions

Structures are locally constructed with ASE and standardized or checked with spglib; rendering uses ASE's headless matplotlib renderer. No Materials Project or COD download is required for this release. There are 48 curated classic prototypes, including 28 derived from pinned AFLOW prototype sources.

Scenes use three tilted crystal directions [100], [110], [111] with an 8-degree tilt and five seeded random directions with zero extra tilt. Q6 uses separate strict, zero-tilt renders. Random directions are sampled as lattice basis coefficients, not uniformly over Cartesian directions in non-cubic cells. Most questions show three default views; Q5 shows two groups of three and Q6 shows one strict view. White backgrounds, element legends, cell edges, bonds, depth ordering, depth fog, and high-contrast highlights provide visual cues.

The cell display includes periodic boundary images, supercell repeats the cell, and environment displays a selected local shell. Atom mappings retain original indices and lattice translations. Q1a counts every displayed image; Q1b accounts for shared boundary atoms. A counted Q1a atom must be at least 50% visible in one default view. Q2 centers and Q3 labels meet the generator's 70% visibility threshold across the eight candidate views.

Bonding and Q2 share a deliberate convention: each structure defines allowed element pairs; for each pair find its shortest nonzero periodic distance d_min; connect distances <= 1.10 * d_min. This is a reproducible benchmark rule, not a universal chemical-bond definition. Nine structures with unsuitable or ambiguous shells are excluded from Q2, leaving 39 eligible base structures.

Q3 uses displayed 3D Euclidean distances without minimum-image wrapping; the distance difference is at least 15% of the larger distance. Half of full-set Q3 questions have depth_trap=true, meaning the first default view reverses the projected versus 3D distance ordering. Other supplied views may resolve it. Q5 has 50 positive and 50 negative full-set pairs. Negatives exchange elements, delete one atom, stretch one lattice axis by 8%, or substitute a same-composition polymorph. Symmetry-equivalent exchanges are rejected. Absolute lattice size counts as part of structure identity; deletion and polymorph changes belong to the atomic-arrangement category. Q6 rejects symmetry-equivalent options using spglib operations.

Evaluation context and limitations

• Small, correlated sample. Mini has only 25 questions per stratum and repeated base structures. Report sample counts and structure-cluster bootstrap intervals. A single structure alone does not identify a between-structure bootstrap interval. No claim is made about unseen structures, experimental microscopy, disorder, defects, or all crystal systems.
• Question and option priors remain. In the accompanying formal run, Claude answered 56% of Q1a without images (14/25; structure-cluster 95% CI 36–76%). Such performance can come from priors or guessing shortcuts rather than counting. Q1b can exploit familiar stoichiometries and conventional-cell knowledge. Balanced answer letters do not eliminate content-based shortcuts.
• Text has more complete information. Coordinates expose hidden positions and periodic-image details. High text performance is not a pure knowledge upper bound or an information-equivalent comparison with occluded projections.
• Simplified chemistry and geometry. These are idealized finite renderings with a fixed neighbor rule and selected color/visibility policies. Conclusions apply to this setup, not every valid chemical coordination definition.
• Third-party endpoints. The completed GPT and Claude runs used Nexfaro and Heju routing, respectively. Returned model IDs are provider-reported metadata, not independent identity or version verification. Attribute results to the tested routes and configurations; no vendor endorsement is implied.
• Kimi is unscored. Its final eight-question smoke test returned five valid responses and three upstream HTTP 502 failures after retries. It was excluded from the formal comparison and ablation. This is missing coverage (N/A), not zero accuracy or evidence about model capability.
• Parsing failures count as wrong. The formal GPT no-image run included 34 invalid outputs. Images must not be added to rescue no-image refusals. API diagnostics and historical pilots are excluded from final accuracy cohorts.
• View-ablation coverage is separate. The compact package omits unused candidates and visibility metadata. Reconstruct complete scenes before making 1/2/4/8-view or visibility-selection comparisons. Shared eight-view endpoints are reused responses, not independent observations.

License, attribution, and citation

Generated benchmark questions and rendered PNGs are released under Creative Commons Attribution 4.0 International, with the full text in LICENSE. Attribute the dataset to MV-CrystalBench contributors (2026) and indicate changes. Companion code is licensed separately under MIT.

AFLOW source material retains its Apache License 2.0 terms. The retained license, NOTICE, selection manifest, and prototype bibliographies are under provenance/aflow/. Source revision: fb7421bb2522888b6a0a4ff91e124c6c65191096 in the AFLOW prototype encyclopedia. The compact package does not include the source CIFs; see provenance/README.md. Benchmark bond policies and question annotations are project additions.

Suggested citation:

@misc{mvcrystalbench2026,
  title = {MV-CrystalBench: Multi-view Crystal Spatial Reasoning Benchmark},
  author = {{MV-CrystalBench contributors}},
  year = {2026},
  url = {https://huggingface.co/datasets/chessnash01/mv-crystalbench},
  note = {716-question full set and 200-question mini subset; local release dated 2026-10-22}
}

AFLOW library references: Mehl et al. (2017), Hicks et al. (2021), and Eckert et al. (2024).