MCC-02: A whole-bean image dataset for four-class cocoa bean quality grading, with a field evaluation set

MCC-02 contains 16,000 laboratory images of whole dried cocoa beans (Theobroma cacao, clone MCC-02) from post-harvest facilities in the Luwu Raya region, South Sulawesi, Indonesia, with 16,013 bounding-box annotations (COCO format) in four quality classes: Fermented, Unfermented, Broken and Moldy. Each bean was photographed whole against a plain white background and then cut; Fermented, Unfermented and Moldy labels were assigned from the cut surface, and Broken from external physical condition, following SNI 2323:2008. The images therefore show the exterior of each bean while the labels reflect its interior, supporting non-destructive prediction of cut-test outcomes. The release also contains:- train/validation/test splits (70/15/15, assigned at image level), with image size, acquisition device and known-issue flags;- a field set of 128 smartphone images (204 beans) captured at working post-harvest facilities on controlled, jute-sack and drying-net backgrounds, with image-level labels. Field beans were not cut: labels were assigned from external appearance, with a few beans from each drying batch cut to confirm the batch status (fermented, unfermented or mouldy), so field labels are less certain than the laboratory labels. This note corrects the statement in README.md v1.0 that the field set was labelled under the same protocol;- field and laboratory prediction records of the twenty model configurations reported in the associated article;- the anchor configuration found by differential evolution;- inter-rater data: three officers re-labelled a stratified sample of 200 images from the photographs (Fleiss' kappa 0.9692 on the 179 images labelled within the four-class scheme; 0.9732 on all 200 with non-standard labels mapped to the SNI classes). The laboratory images are split into 32 archives (8 per class); extracting all of them into one folder rebuilds the full set. See README.md for file structure, label definitions, acquisition devices and known issues. Code: https://github.com/ramsis235/mcc02-cocoa-bean-grading