Datasets:

---

privbench

/

data

Like 1

Follow

privbench 3

Adds the imputation_accuracy key to 20 attribute-inference ground-truth files. Everything else in each file is byte-for-byte
the current org content (verified key-by-key); the JSONs only gain one number.

attacks/attribute/scoring/score_attribute.py (privacy-arena main, since 137ed40 on 2026-07-16) uses imputation_accuracy as the
reward baseline for graybox/label attribute inference and falls back to base_rate only when the key is missing. The July run
that computed these values wrote them to the local benchmark_data store and to dwahdany/privacy-arena-data, never to this
dataset. All August agent evaluations were scored with these files.

Not included: the 4 adult_tabpfn_* ground truths (their local copies never received the key, and 3 of them differ from the org
values; left for the task owner) and the models/mia/texas_tabpfn_* files, where this dataset is authoritative.