Steel surface defect detection benchmark with test-set uncertainty quantification: data and code (NEU-DET, GC10-DET; YOLOv8, YOLO11, YOLO12, RT-DETR)

Data and code for "How much of a per-class gain is noise? A controlled multi-seed benchmark of YOLOv8, YOLO11, YOLO12 and RT-DETR for steel surface defect detection with test-set uncertainty quantification" (Jeong and Song, submitted to PLOS ONE). Contents: split manifests of the stratified 70/15/15 re-splits of NEU-DET (1,799 images) and GC10-DET (2,292 images); per-image evaluation records (prediction confidence, class and IoU-0.5 true-positive flag; ground-truth classes) for all 41 training runs of the study and 3 sensitivity runs; per-epoch training logs and arguments of every run; result tables reported by the framework; deployment measurements (PyTorch, TensorRT FP16, ONNX); confusion counts; data-set statistics; the training harness for the class-imbalance remedies; and the analysis code (NumPy/pandas/SciPy/Matplotlib, no GPU needed) that reproduces every number, table and figure of the article from these files. See README.md. Images and annotations are not included; they are available from the public mirrors cited in the README (CC BY 4.0).