Datasets:

---

privbench

/

data

Like 1

Follow

privbench 3

Adds the 25 files that so far existed only in the personal mirror dwahdany/privacy-arena-data, which is about to be deleted.
Every file describes the model that is currently in this dataset:

• 19 utility caches (.verify_cache.json, .lm_eval_cache.json) whose recorded model hash (sha256 of the first 64 KB, as computed by
  attacks/*/verify_utility.py) equals the model file in the same folder here. The dataset already keeps 7 such caches.
• 6 ViT-B/16 side artifacts (full_pool_scores.npz, training_history.json, training_summary.json for fairface and fitzpatrick17k).
  They are the input of the two ViT rows of the paper's target-utility table.

Deliberately NOT included: 14 caches written for models this dataset has since replaced (folktables attribute models, OLMo dataset-
inference shadows, adult/folktables/texas TabPFN MIA targets). They are archived in the code repo instead, so nothing here describes a
superseded model. All other mirror-only content is already in this dataset's main branch or history.