GeoComb public data and model package for frame-level honey bee comb analysis

Supporting data and model package for GeoComb, a deep-learning-assisted and geometry-aware workflow for frame-level honey bee comb analysis. The archive contains 53 in-house full-frame images with annotations, 1,065 corrected ROI annotation records, deduplicated 159-image ROI test results, portable manifests for 1,241,053 classifier crops, four final model-weight files, provenance records, and SHA-256 checksums. One exact train–test ROI duplicate was excluded from corrected test evaluation. Raw DeepBee and FAIRHiveFrames-1K images are not redistributed and must be obtained from their original sources.