Diffusion Generator Weights and Generated Pools for "When Validation-Gated Selection of Diffusion Augmentation Fails to Generalize: A Retrospective Audit of Rare Seagate KQI Classification"
Abstract
This data set contains the fitted generator weights and the generated positive pools of the rebuilt candidate archive described in the manuscript "When Validation-Gated Selection of Diffusion Augmentation Fails to Generalize: A Retrospective Audit of Rare Seagate KQI Classification" by Tae-Yoon Kim, Young-Shin Han, Jong-Sik Lee, and Bo-Seung Kwon (Inha University; under review at Electronics, MDPI). The candidate archive was rebuilt for all 11 Seagate time-series-2 KQI endpoints and five repeats with a corrected mixed-type diffusion generator (Gaussian diffusion for continuous features and two-category multinomial diffusion for binary features). The data are split into one zip file per endpoint (rebuilt_archive_endpoint_00.zip to rebuilt_archive_endpoint_10.zip) and a small zip with the run manifest and log (rebuilt_archive_run_metadata.zip). Unzipped into one folder, the twelve files recreate the folder regen_archive_v1_run1/, with one subfolder per endpoint-repeat case (55 cases). Each subfolder holds the weights of the two generators (float16, exponential moving average), their input transforms, the generated pools, the case record with the generator diagnostics, and the validation and test metrics of every candidate for LightGBM and XGBoost. SHA256SUMS.txt lists the SHA-256 of every zip file and of every file inside them, and README.txt describes the layout and how to use the files with the code. The code, configuration, protocol, and lightweight outputs that use these files are archived in the companion software record (version 2.0: https://doi.org/10.5281/zenodo.23233292; GitHub: https://github.com/tyoonkk/seagate-kqi-diffusion-audit, folder revision_2026_10/). Its EXCLUDED_HEAVY_FILES.json lists the weight and pool files with the same SHA-256 hashes. The Seagate data are not included; they are available from https://github.com/Seagate/softsensing_data under the Apache-2.0 license.
