QuasiLeak-Bench

Paper · Code · Leaderboard

QuasiLeak-Bench is a benchmark for policy-conditioned quasi-evidence group recovery. Messages often contain cues that are weak alone but together reveal a private attribute of the sender, and we call such a set of cues a quasi-evidence group. Each document comes with the sender's profile, the communication context, and a disclosure policy that marks each of eight attributes HIDE, KEEP, or NEUTRAL. A detector returns the groups that reveal a HIDE attribute, each as two to four verbatim spans bound to that attribute.

Configurations

config	split	documents	use
qlb (default)	train / validation / test	3,500 / 500 / 1,000	QLB-Test is the leaderboard set
full	full	5,000	prompted systems, as in the paper's main comparison
subset_non_revealing	full	1,618	QLB-SNR (Subset Non-Revealing), Full documents in which no smaller subset of any HIDE group reveals its attribute
policy_shift	full	500 pairs	QLB-PS, the paper's policy study: documents under their released and a reassigned policy
sft_policy_shift	test	500 pairs	the fine-tuning study's policy pairs, all in Test

from datasets import load_dataset

test = load_dataset("HenryZYF/QuasiLeak-Bench", "qlb", split="test", revision="v1.0.1")

The split is by profile, 350 / 50 / 100 profiles, so no Test profile appears in Train or Validation. Test lies inside Full, so a system trained on any Test document is not eligible for the leaderboard. QLB-PS and QLB-SNR are views of Full. For QLB-SNR, every smaller subset of each HIDE trio and quad is checked with the benchmark's verifier.

Versions

This is v1.0.1, the version on the leaderboard. The code repository keeps v1.0.0 and the changelog.

Fields

field
sample_id	document id
document	message text
profile	profile_id and the eight attributes age, gender, nationality, education, occupation, income, relationship, health
context	context_id, name, recipient, purpose, description, length (words)
policy	hide, keep, neutral, a partition of the eight attributes
gold.groups[]	group_id, attribute, cue_ids (2 to 4), secondary_attributes, distance, verification
gold.spans[]	cues: span_id, group_id, text, start, end
gold.distractors[]	distractor_id, text, start, end

Offsets are half-open code-point indices into document. A policy-shift row carries policy_original and policy_reassigned in place of policy, the group_ids each policy hides (targets_original, targets_reassigned), and the attributes whose HIDE status changes (newly_hidden, newly_permitted).

Scoring

The scorer and evaluation prompts are in the code repository. The primary metric is P-GF1, the policy group F1, at token IoU 0.75. Scores include the false-alarm rate on documents with no HIDE group, and QLB-PS adds the paper's Suppress and Surface measures.

Intended use

QuasiLeak-Bench is for evaluating and developing sender-side privacy detectors. The profiles are synthetic and describe no real person. Do not use the data to infer private attributes of real people. 300 profiles are adapted from SynthPAI (Yukhymenko et al., 2024) under the same license.

Citation

@article{zhu2026quasileak,
  title   = {QuasiLeak-Bench: Evaluating LLM-Based Privacy Detectors on
             Policy-Conditioned Quasi-Evidence Group Recovery},
  author  = {Zhu, Yufan and Jiang, Yangfan and Fei, Yang and Hou, Shizheng and
             Mohapatra, Shubhankar and Xiao, Xiaokui},
  journal = {arXiv preprint arXiv:2610.21333},
  year    = {2026}
}