Аннотация
Field-deployable plant-health vision systems require datasets that preserve natural variation in illumination, scale, occlusion, and background clutter. This Data Report describes a fixed public release of 500 RGB JPEG images of cucumber leaves collected at three protected-cultivation facilities in Kazakhstan. The collection contains 250 healthy-leaf images and 250 affected-leaf images. Decoder-level inspection confirmed three native image dimensions: 3840 × 2160 pixels (387 images), 4608 × 3456 (112), and 4000 × 2252 (one). The affected subset is paired with 250 text files containing 2,705 rectangular annotations in normalized YOLO center-coordinate format: 1,741 class-0 boxes identify visible non-green color and 964 class-1 boxes identify visible tissue damage. Class 1 includes visible tissue deterioration regardless of whether the apparent cause is disease-associated, mechanical, pest-related, senescent, or otherwise uncertain. The two phenotype labels are operationally non-exclusive: 215 affected images contain both classes, and 116 contain at least one positive-area cross-class box overlap. These labels describe observable phenotypes and do not provide pathogen, disease-species, cultivar, or causal diagnoses. The release supplies bounding boxes, not pixel-level masks; it therefore supports image classification, object detection, and rectangular region localization, but not conventional supervised segmentation. A complete decoder, structure, coordinate, exact-hash, filename, and coarse perceptual-hash audit confirmed one-to-one affected image-label pairing and valid five-field rows, while identifying 42 boundary-crossing boxes in 36 files, 19 byte-identical pairs, and three additional high-similarity candidate pairs. The intentionally balanced final folders are a curated sample and must not be interpreted as natural prevalence. The dataset is publicly available in Mendeley Data, version 1, under CC BY 4.0 at https://doi.org/10.17632/mb3vnhpk8z.1.