Commit History
-
Commit: Updated README.md
Verified signature1. Align the checkpoint inventory, corrected multi-seed results, terminology, and checkpoint-selection guidance with the new 600-checkpoint release.a8dd6fe -
Update corrected experiment checkpoints (part 3)
Verified signatureReplace pixel-only, low-frequency-only, and mixed-domain checkpoints for seeds 42-46; remove obsolete baseline replication checkpoints.e552740 -
Update corrected experiment checkpoints (part 2)
Verified signatureReplace pixel-only, low-frequency-only, and mixed-domain checkpoints for seeds 42-46; remove obsolete baseline replication checkpoints.6f3af8b -
Update corrected experiment checkpoints
Verified signatureReplace pixel-only, low-frequency-only, and mixed-domain checkpoints for seeds 42-46; remove obsolete baseline replication checkpoints.d2fa73c
-
Commit: Updated README.md
Verified signature1. Documented 5-seed mixed-domain checkpoints (seed-42 through seed-46), expanded the results table with union robustness and peak/final metrics, and added usage examples for downloading mixed-domain weights (epoch_85.pt).9cb4f0e -
Upload mixed-domain seed-46 checkpoints (epochs 5-200)
Verified signatureac5f4ea -
Upload mixed-domain seed-45 checkpoints (epochs 5-200)
Verified signature8dea342 -
Upload mixed-domain seed-44 checkpoints (epochs 5-200)
Verified signature465fa57 -
Upload mixed-domain seed-43 checkpoints (epochs 5-200)
Verified signature7a509a9 -
Upload mixed-domain seed-42 checkpoints (epochs 5-200)
Verified signature3365a3b
-
Commit: Update README.md
Verified signature1. Added multi-seed low-frequency evaluation benchmarks (comparing clean, pixel-PGD, and low-frequency-PGD accuracies), training dynamics observations, and checkpoint usage recommendations to the results section.e27aed7 -
Commit: Update README.md
Verified signature1. Updated repository structure, training specifications, evaluation benchmark results, and checkpoint loading examples to include the 5 newly released multi-seed pixel-only runs (seed-42 through seed-46).f5bff66 -
Upload folder using huggingface_hub
Verified signatureef433d3 -
Upload folder using huggingface_hub
Verified signatureccb6b79 -
Upload folder using huggingface_hub
Verified signature19dd90a -
Upload folder using huggingface_hub
Verified signatured2d6a72 -
Upload folder using huggingface_hub
Verified signaturef14880f
-
Commit: Update README.md
Verified signature1. Add frequency-domain and discrete-cosine-transform metadata tags. 2. Document repository layout with baseline-replication/ and low-frequency-only/ seed directories. 3. Expand model details to cover 5-seed low-frequency DCT-masked PGD-10 runs. 4. Update hf_hub_download code snippets to include subfolder paths. 5. Refine limitations section to reflect the multi-condition experiment suite.2869f36 -
Remove diagnostic folder
Verified signatureec71911 -
Remove loose root-level checkpoint files (moved to baseline-replication/)
Verified signature34e6ba7 -
Upload folder using huggingface_hub
Verified signature1fdf2d5 -
Upload folder using huggingface_hub
Verified signature6ad021f -
Upload folder using huggingface_hub
Verified signaturea317397 -
Upload folder using huggingface_hub
Verified signature4188392 -
Upload folder using huggingface_hub
Verified signature94d4dc6 -
Upload folder using huggingface_hub
Verified signature85eb60f
-
Commit: Delete .DS_Store
Verified signature1. Deleting accidentally uploaded .DS_Store file from macOS.b949202 -
Commit: Create README.md
Verified signature1. 40 PreActResNet-18 checkpoints trained on CIFAR-10 with pixel-space PGD-10 adversarial training. 2. Covers a complete 200-epoch run, with checkpoints saved every 5 epochs. 3. Intended for reproducing and analyzing robust overfitting: robust test accuracy peaks and then declines during continued adversarial training. 4. The strongest measured checkpoint is `epoch_105.pt`, reaching 45.91% robust accuracy under PGD-20 evaluation. 5. `epoch_200.pt` is the final checkpoint; its PGD-20 robust accuracy is 36.18%. 6. Training uses an \(L_\infty\) perturbation budget of 8/255, a PGD step size of 2/255, and random seed 42. 7. Full training code, evaluation scripts, and results are available at [ItsKaiwenDu/Robust-Overfitting](https://github.com/ItsKaiwenDu/Robust-Overfitting). 8. These weights are for research and reproduction; they are not a claim of state-of-the-art robustness.dad6b25
-
Upload folder using huggingface_hub
Verified signaturec939579 -
initial commit
Verified signature25ec4a7