max6616/SiD3M-test

SiDIT test data

This release contains 9,969 recovered cross-library test scans for the SiDIT base model described in Joint reflection-family and crystal-property prediction from single-crystal diffraction point clouds. The software and benchmark result support and model parameters use one fixed configuration and checkpoint. The dataset reference is jac-2026-10-12.

Files and use

• ood-test.tar.gz: verified OOD test-input records in JSONL format, with source record bytes preserved.
• ood-test-manifest.jsonl.gz: filename, source-record SHA256, CIF identifier, raw/effective point counts and benchmark label checks.
• source-cif-provenance.json.gz: exact-byte matches from source CIF identifiers to COD/NOMAD filenames and CIF hashes.
• RECOVERY_SCOPE.json: evaluation and raw-input scopes, source identities and exclusions.
• CHECKSUMS.json: file sizes and SHA256 digests.

Extract with tar -xzf ood-test.tar.gz. Each JSONL file stores one simulated scan with input_sequence rows [rotation_angle_degrees, normalized_detector_x, normalized_detector_y, intensity], canonical reflection-family labels rows [h,k,l], and crystallographic reference metadata. The component labels cover [-10, 10]. Prediction examples use the SiDIT loader and deterministic 0.0001 voxel preprocessing; raw and evaluated point counts can therefore differ.

Benchmark support

The OOD evaluation contains 9,970 scans. The raw-input archive contains 9,969 matching scans and excludes one input that disagrees with its cached point count, together with 11 files outside the evaluation cohort. The result support preserves the 9,970-scan statistics. RECOVERY_SCOPE.json lists the exclusions. Recovered inputs are checked by filename, symmetry label and effective point count, with SHA256 digests recorded in the manifest; original input-file hashes are unavailable in the evaluation cache.

The source pool comprises COD and NOMAD CIFs. The input subset refers to 4,326 source CIF identifiers and 3,367 distinct CIF byte hashes. These counts describe source-file identities. Physical-structure deduplication and complete per-candidate Materials Project overlap-screening records are outside the documented checks.

The benchmark records contain predictions and reflection counts for 14,990 main-test scans and 9,970 OOD scans, supporting recalculation of the reported summaries. The original 14,990 main-test raw inputs have not been recovered. Training and validation inputs are excluded. This dataset uses tar/JSONL format, and the automatic Dataset Viewer is disabled.

The interpretation notes define point thinning, angular clipping and coordinate corruption, including the full-scan intensity descriptor supplied to symmetry prediction in those tests. Evaluation arrays retain their own order and metric denominators.

Attribution and license

Derived simulated test records and release metadata are distributed under CC BY 4.0. COD source data are dedicated to CC0/public domain. NOMAD open content uses CC BY 4.0; source filenames are retained in the provenance mapping. Source crystal-coordinate CIF files are outside this release.

Authors: Zhao Zhang, Zheng Dong, Zhi Geng, Xinlong Dong, Yi Zhang and Gaoqi He. Corresponding author: Gaoqi He, East China Normal University. See CITATION.cff for dataset citation metadata.