Datasets:
add README (disable viewer, document structure)
Browse files
README.md
ADDED
|
@@ -0,0 +1,49 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
viewer: false
|
| 3 |
+
tags:
|
| 4 |
+
- adversarial-robustness
|
| 5 |
+
- image-classification
|
| 6 |
+
- robustness-benchmark
|
| 7 |
+
---
|
| 8 |
+
|
| 9 |
+
# RobustGenBench-sample
|
| 10 |
+
|
| 11 |
+
This is a **stratified sample** of the full [RobustGenBench](https://huggingface.co/datasets/legolasflagstaff/RobustGenBench) dataset (22 GB), created for NeurIPS dataset submission review.
|
| 12 |
+
|
| 13 |
+
## Structure
|
| 14 |
+
|
| 15 |
+
Each subfolder contains **2 images per class** extracted from the corresponding archive in the full dataset, plus `test/labels.csv` and `metadata.json`.
|
| 16 |
+
|
| 17 |
+
```
|
| 18 |
+
clean/
|
| 19 |
+
caltech101/ ← 2 imgs/class × 101 classes = 202 images
|
| 20 |
+
fgvc-aircraft-2013b/
|
| 21 |
+
flowers-102/
|
| 22 |
+
oxford-iiit-pet/
|
| 23 |
+
stanford_cars/
|
| 24 |
+
uc-merced-land-use-dataset/
|
| 25 |
+
|
| 26 |
+
adversarial/
|
| 27 |
+
common/common_severity3/<dataset>/
|
| 28 |
+
random/linf_eps30_random_uniform/<dataset>/
|
| 29 |
+
zeroshot_clip_vitb16_laion2b/<threat_model>/<dataset>/
|
| 30 |
+
zeroshot_clip_vith14_laion2b/<threat_model>/<dataset>/
|
| 31 |
+
zeroshot_metaclip_vith14_fullcc2_5b/<threat_model>/<dataset>/
|
| 32 |
+
zeroshot_siglip2_base_patch16_224/<threat_model>/<dataset>/
|
| 33 |
+
zeroshot_siglip2_so400m_patch14_384/<threat_model>/<dataset>/
|
| 34 |
+
zeroshot_siglip2_so400m_patch16_naflex/<threat_model>/<dataset>/
|
| 35 |
+
zeroshot_siglip2_so400m_patch16_naflex_patchify/<threat_model>/<dataset>/
|
| 36 |
+
```
|
| 37 |
+
|
| 38 |
+
Each leaf folder has the same internal layout:
|
| 39 |
+
- `test/labels.csv` — `filename,label` mapping (integer class indices)
|
| 40 |
+
- `test/NNNNN.png` — flat-numbered PNG images (same filenames across clean & adversarial)
|
| 41 |
+
- `metadata.json` — split counts (clean archives only)
|
| 42 |
+
|
| 43 |
+
Class name mappings are in `class_names/<dataset>.json`.
|
| 44 |
+
|
| 45 |
+
See **[SAMPLE.md](SAMPLE.md)** for full details on sampling methodology.
|
| 46 |
+
|
| 47 |
+
## Full dataset
|
| 48 |
+
|
| 49 |
+
👉 https://huggingface.co/datasets/legolasflagstaff/RobustGenBench
|