Verified adversarial examples for efficientnetv2l (torchvision/EfficientNetV2LWeights.IMAGENET1KV1), produced by the Perturb network. Each row is one clean image together with all of its (L∞ ≤ 0.03 in [0,1] pixel scale) yet change the model's top-1 prediction. This dataset grows continuously. New rows are appended as the network produces them and uploaded in batches; re-run loaddataset (or pass downloadmode="forceredownload") to pick up the latest. Adversarial training / fine-tuning of image classifiers, and evaluation of robustness against small L∞ perturbations. For each row, (image, label) is a clean training pair and every element of adversarial is a hard example that should be…