TTA-CIFAR-100-C

Mirror of CIFAR-100-C (Hendrycks & Dietterich, ICLR 2019) with a revision pin for reproducible test-time adaptation evaluation.

• Upstream: Zenodo record 3555552
• License: CC BY 4.0 (matches upstream)
• Sibling: TTA-CIFAR-10-C
• Maintained as part of: TTA-Evaluation-Harness

Citation

@inproceedings{hendrycks2019benchmarking,
  title={Benchmarking Neural Network Robustness to Common Corruptions and Perturbations},
  author={Hendrycks, Dan and Dietterich, Thomas},
  booktitle={ICLR}, year={2019}
}

Structure

• 19 configs: 15 standard corruptions + 4 held-out corruptions (gaussian_blur, saturate, spatter, speckle_noise). The TTA community typically reports only the standard 15.
• 5 splits per config: severity_1 through severity_5, 10 000 images each.
• Same 10 000 CIFAR-100 test labels are reused under every corruption / severity.

Usage

from datasets import load_dataset

ds = load_dataset("WNJXYK/TTA-CIFAR-100-C",
                  name="gaussian_noise",
                  split="severity_5",
                  revision="v1.0")

Provenance

Built by scripts/publish_cifar100c.py in the TTA-Evaluation-Harness repo from the upstream .npy files (no pixel-level re-generation). Individual per-corruption files have sha256 recorded alongside.