Data and scripts for "Structured Pruning of YOLO26n for Insulator-Defect Inspection: A Reproducible Accuracy–Complexity Study"
Description
Evaluation outputs and reproduction scripts for a three-seed, budget-matched comparison of structured channel-pruning criteria (L1, FPGM, LAMP) on YOLO26n for insulator-defect inspection. Contains: the master results table (three-seed means +/- sd, per-class AP, parameters, GFLOPs, pure-forward FPS, cross-domain insulator AP); per-run Ultralytics evaluation results on the Insulator-Defect test split for 27 training runs across seeds 42/43/44; per-class COCO evaluations (areaRng [0,1e9]); zero-shot cross-domain results on CPLID and InsPLAD-det; knowledge-distillation ablation; locally calibrated GFLOPs measurements; figure data; the source-independent re-partition used for the headline claims (manifest of the 400 source photographs with their four augmented variants, per-split file lists, seed 42, plus the dataset config) with the per-epoch results of the three arms re-trained on it; and the evaluation, statistics and figure scripts, including the split-contamination probe and a 34-assertion script that recomputes every derived number quoted in the manuscript. The three benchmark datasets (Insulator-Defect, CPLID, InsPLAD-det) are publicly available from their original publishers and are not included. No model weights are included. See README.md for the run-name mapping and file inventory.