Skip to content
DSML.KZ
Publications

Dataset · Zenodo · October 3, 2026

Data and code for: Public-image lineage defects alter model rankings without a uniform clean-selection gain

Xin Li, Bojian Guo, A.Dzh. Kartanova

Asel Kartanova
Open via DOI

Abstract

Derived manifests, deterministic lineage evidence, split assignments, protocol locks and amendments, fixed model/checkpoint identifiers, executable analysis and figure code, row-level predictions, aggregate results, validation receipts and checksums for a common-anchor audit of public image benchmark lineage defects. The release supports independent recomputation from archived predictions and regeneration of publication figures. A fresh source-image training replay is not claimed as a validated one-command procedure; historical environment-dependent locks are retained transparently. Third-party source images and regenerable feature arrays are not redistributed. Version-pinned source DOIs, licences, hashes and rebuild instructions are provided. The primary clean-selection regret was +0.011027 (95% interval +0.002917 to +0.019553), below the prespecified +0.02 criterion. The outcome is inconclusive under that criterion; the archive does not establish a uniform gain from clean model selection. The CC BY 4.0 licence covers the authors’ own code, derived outputs and documentation. Third-party source metadata and datasets remain subject to their original terms. The archive excludes raw third-party images, manuscript submission administration, signatures, journal forms and full-text source snapshots. Download instructions: Download multipart01.zip through multipart05.zip and extract ALL five ordinary ZIP archives into the same destination directory. They contain disjoint files under the same Public_Image_Lineage_Audit_v1.0 folder. Do not concatenate the ZIP files. See DOWNLOAD_AND_REPRODUCE.txt for verification and reproduction commands.

Scientific Computing and Data ManagementMachine Learning and Data ClassificationResearch Data Management Practicesmodel rankingreproducibility auditimage classification models

More by Asel Kartanova