Leak-Free Meat Freshness Benchmark: A Cleaned Nine-Class Fixed-Split Image Dataset

Abstract

This record provides a cleaned, leak-free benchmark derived from version 1 of the Mendeley Data collection “A Comprehensive Image Dataset for Accurate Authentication of Meat Species and Hourly Freshness Using Artificial Intelligence in Food Safety” (https://doi.org/10.17632/4tj9t3n6vj.1). The benchmark contains 1,849 original JPEG images organized into nine meat species and freshness classes, with a fixed train/validation/test split of 1,290/277/282 images. Offline-augmented images were excluded, and exact byte-identical files were audited using MD5. The byte-identical 36 h and 48 h Beef folders were represented once as “36+ hr Beef,” whereas the distinct 36 h and 48 h Mutton classes were retained separately. The package includes the cleaned images, a per-image MD5 manifest, the fixed split assignments, a machine-readable verification report, source attribution, and a detailed description of all benchmark modifications. The resulting benchmark contains no exact-file overlap between the training, validation, and test partitions. Audit, deterministic reconstruction, leakage verification, model training, statistical evaluation code, and the reported four-model five-seed results are available at https://github.com/AhmetCannnn/leak-free-meat-freshness-benchmark. The corresponding frozen software release is version v1.0.0 at commit dacaf986ab2052577ea827c1fb5d541b3c98eadc.