README.md

metadata

viewer: false
tags:
  - adversarial-robustness
  - image-classification
  - robustness-benchmark

RobustGenBench-sample

This is a stratified sample of the full RobustGenBench dataset (22 GB), created for NeurIPS dataset submission review.

Structure

Each subfolder contains 2 images per class extracted from the corresponding archive in the full dataset, plus test/labels.csv and metadata.json.

clean/
  caltech101/          ← 2 imgs/class × 101 classes = 202 images
  fgvc-aircraft-2013b/
  flowers-102/
  oxford-iiit-pet/
  stanford_cars/
  uc-merced-land-use-dataset/

adversarial/
  common/common_severity3/<dataset>/
  random/linf_eps30_random_uniform/<dataset>/
  zeroshot_clip_vitb16_laion2b/<threat_model>/<dataset>/
  zeroshot_clip_vith14_laion2b/<threat_model>/<dataset>/
  zeroshot_metaclip_vith14_fullcc2_5b/<threat_model>/<dataset>/
  zeroshot_siglip2_base_patch16_224/<threat_model>/<dataset>/
  zeroshot_siglip2_so400m_patch14_384/<threat_model>/<dataset>/
  zeroshot_siglip2_so400m_patch16_naflex/<threat_model>/<dataset>/
  zeroshot_siglip2_so400m_patch16_naflex_patchify/<threat_model>/<dataset>/

Each leaf folder has the same internal layout:

• test/labels.csv — filename,label mapping (integer class indices)
• test/NNNNN.png — flat-numbered PNG images (same filenames across clean & adversarial)
• metadata.json — split counts (clean archives only)

Class name mappings are in class_names/<dataset>.json.

See SAMPLE.md for full details on sampling methodology.

Full dataset

👉 https://huggingface.co/datasets/legolasflagstaff/RobustGenBench