RuralCraftPattern: Dataset, Experimental Results, and Source Provenance for Fine-Grained Traditional Pattern Recognition

This deposit contains the final RuralCraftPattern dataset, experimental results, and corresponding source-provenance materials for fine-grained recognition of traditional decorative patterns under limited-data conditions. The dataset contains 461 processed images in eight classes, with train, validation, and test splits of 320, 48, and 93 images, respectively. The package includes sample-level metadata, per-sample predictions, run-level outputs, main comparison results, ablation experiments, robustness evaluations, statistical test results, checksums, and source-acquisition records. The source_provenance directory contains the confirmed source-acquisition registry corresponding to the image collection. The registry contains 1,483 acquisition records, including 930 records retained during the acquisition stage and 553 additional candidate records. It preserves original local paths, image URLs, object URLs, source object identifiers, source types, license statements, copyright status, and download dates. The source registry retains six acquisition-stage labels, while the current processed dataset is organized into eight task-specific classes. The current RCP-* filenames were normalized after image selection and processing. Because stable original row identifiers were not retained in the supplied current filenames, the source registry documents provenance at the acquisition-collection level rather than through an artificially generated one-row-per-image mapping. Across three random seeds, MSF-SViT achieved a Top-1 accuracy of 0.7742 ± 0.0323, a Macro-F1 score of 0.7770 ± 0.0263, and a balanced accuracy of 0.7739 ± 0.0262. This final package was assembled by merging and organizing the supplied images, metadata, experimental results, and source-provenance records. Model outputs and experimental values were not recomputed or modified during packaging.