max6616/SiD3M-test

SiDIT historical test data

This release contains 9,969 recovered out-of-library test scans for SiDIT, the point-cloud method described in Joint reflection-family and crystal-property prediction from single-crystal diffraction point clouds. The core software and compact retained-result support are available at max6616/SiDIT, and the tensor-only model at max6616/SiDIT on Hugging Face.

Files and use

• ood-test.tar.gz: only the verified historical OOD test-input subset, with original JSONL record bytes preserved.
• ood-test-manifest.jsonl.gz: filename, original-record SHA256, CIF identifier, raw/effective point counts and cache label checks.
• source-cif-provenance.json.gz: exact-byte matches from balanced CIF identifiers to retained COD/NOMAD source filenames and source CIF hashes.
• RECOVERY_SCOPE.json: exact release, original-cache and recovery scopes, including every exclusion.
• 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 holohedry labels rows [h,k,l], and crystallographic reference metadata. The original source labels cover [-10, 10]. Prediction examples use the released SiDIT loader and deterministic 0.0001 voxel preprocessing; raw and evaluated point counts can therefore differ.

Relation to the retained paper results

The original OOD evaluation contains 9,970 scans. The later backup contains 11 appended files absent from that evaluation and one cache-named file whose input disagrees with its cached point count. The public input archive excludes these files and contains 9,969 scans; the retained 9,970-scan result arrays remain unchanged in the compact support release. The excluded file is listed in RECOVERY_SCOPE.json. Historical cache records did not retain original input-file hashes; the recovered inputs have now been hashed and checked against cache filenames, symmetry labels and effective point counts.

The source pool was built from COD and NOMAD CIFs. The released input subset refers to 4,326 source CIF identifiers, representing 3,367 distinct CIF byte hashes. These are source-identifier and file-content counts; they do not establish the number of unique physical structures. Original MP-overlap screening logic is retained in the project history, but complete per-candidate screening logs are unavailable. The release supports inspection of the historical simulated benchmark and inference with the published model.

This repository excludes training and validation inputs. The historical in-distribution main-test roots are documented independently; this OOD archive must not be substituted for them. This is a tar/JSONL research-data release; the automatic Dataset Viewer is disabled.

Attribution and license

Derived simulated test records and release metadata are distributed under CC BY 4.0. COD source data are dedicated toCC0/public domain. NOMAD open content usesCC BY 4.0; original source filenames are retained in the provenance mapping. Source crystal-coordinate CIF files are not included.

Authors: Zhao Zhang, Zheng Dong, Zhi Geng, Xinlong Dong, Yi Zhang and Gaoqi He. Corresponding author: Gaoqi He, East China Normal University.

The fixed software/model release for this data is v0.1.1. The separate retained-result support preserves original array order and metric denominators. Its interpretation notes distinguish point thinning, angular clipping and noisy-coordinate perturbations, including the full-scan intensity descriptors retained in those historical tests.

The public GitHub support contains retained prediction/count records for 14,990 main-test scans and 9,970 OOD scans, sufficient to recalculate the published summaries. This dataset release contains recovered raw input records for 9,969 OOD scans. The original 14,990 main-test input scans have not been recovered from the nominated archives; complete replay of that original input cohort is therefore unavailable.