TUPLE-Bench

Auditing privacy–utility trade-offs in typed decision outputs.

Paper (arXiv) · Code and evaluation APIs · Full dataset ZIP · Reproduction guide

Maintained by Qiaoyuan Zheng.

Dataset summary

TUPLE-Bench studies whether ordinary decision fields reveal a protected fine mechanism beyond an authorized coarse category. A recipient may learn NETWORK, for example, without learning whether the cause was a stale resolver entry or a stale overlay route.

The benchmark contains 720 matched secret pairs / 1,440 worlds across delivery operations, equipment service, and IT operations. Paired worlds share public context, recipient permissions, coarse truth, action costs, and permitted checks; their fine mechanism and private evidence differ.

Five release views are derived from the same saved response: no output, choice, choice+confidence, full probabilities, and full native. Privacy is held-out binary secret inference conditioned on the authorized category; utility is realized routing, review, and clarification cost.

Split	Pairs	Worlds	Worlds per domain
Train	432	864	288
Validation	144	288	96
Test	144	288	96

Each domain has four coarse categories and sixteen protected mechanisms. Every mechanism has 18/6/6 train/validation/test worlds. Both worlds of a source family remain in one split; the full ZIP includes four fixed grouped training folds. These are the original frozen assignments.

Files

File	Contents
tuple-bench-dataset.zip	Complete canonical evaluation database and final native responses
train.jsonl, validation.jsonl, test.jsonl	Reversible viewer projection with the exact frozen split
SHA256SUMS.txt	Checksums for the distributed files

The full ZIP contains tuple-bench/data/ and its dataset checksum manifest:

• 1,440 nested world records and 720 paired public specifications;
• domain ontologies, recipient contracts, cost/check specifications, and structural certificates;
• SQLite mirror with worlds, pairs, metadata, public_cases, and model_inputs;
• final coarse native responses for TypeSafe, OpenJev, JevK5, Clef, Clef-flash, Decider, Sage, d1, and Mercury Decide;
• separate nine-model refusal-probe responses;
• recorded model versions, training folds, and retained review issues.

The canonical dataset ZIP is 6,057,682 bytes. SHA-256:

27abf9c9a77efb6b848ffd3ead16ea5577ff8ee91b88993a890f1980d7e38fdd

The dataset is identical to the original v1.0.0 evaluation archive. Generation code is not included. Evaluation software and frozen fitted models are available from GitHub.

Load the viewer tables

import json
from datasets import load_dataset

dataset = load_dataset("Qiaoyuan/TUPLE-Bench", token=False)
assert [len(dataset[s]) for s in ("train", "validation", "test")] == [864, 288, 288]
row = dataset["train"][0]
model_view = json.loads(row["model_view_json"])
consumer_view = json.loads(row["consumer_view_json"])
evaluator_only = json.loads(row["evaluator_only_json"])

The nested views are JSON strings so all three domains share one stable Arrow schema. Parsing those strings recovers the original objects exactly.

Column	Meaning and access boundary
schema_version, case_id, domain_id, contract_id	Original schema, world, and contract identifiers
pair_id, source_family_id, rendering_family_id	Grouping/provenance; never prediction features
split	Original split
authorized_category	Gold coarse T; evaluator-only when consumers act
protected_mechanism	Gold fine S; training/evaluation label
secret_hypotheses	Registered mechanism alternatives for the privacy audit
model_view_json	Public contract and state, private diagnostic evidence, coarse question
consumer_view_json	Public recipient context, candidates, downstream context
evaluator_only_json	Complete original labels, grouping, split, check answers, and provenance

The target model receives only model_view; consumers receive only public context and the chosen release. The attacker additionally knows authorized T and the registered hypotheses, but not private evidence, gold S, case/family IDs, or the paired response at test time.

Download the complete evaluation package

from zipfile import ZipFile
from huggingface_hub import hf_hub_download

archive = hf_hub_download(
    repo_id="Qiaoyuan/TUPLE-Bench", repo_type="dataset",
    filename="tuple-bench-dataset.zip", token=False,
)
with ZipFile(archive) as zipped:
    zipped.extractall("evaluation-release")

For the standalone evaluator, also download the software ZIP and extract it into the same evaluation-release/ parent. Both archives merge into evaluation-release/tuple-bench/. Follow the README there to install and run the evaluation. The GitHub checkout already includes the canonical data and fitted artifacts.

The viewer tables alone do not contain the saved model responses needed to reproduce reported evaluations. Use the full archive for that purpose.

Quality, provenance, and limitations

All cases are AI-authored hypothetical incidents, not real customer records. Structural checks verify shared public fields, coarse truth, costs, and check laws; they do not prove semantic validity or equal informativeness of private narratives.

The frozen development dataset retains three B-review uncertain pairs, one invalid quote record, two related six-pair groups, and other adverse cross-family relation evidence in data/review_issues.json. Acceptance for evaluation is distinct from all-judge approval. Source-family isolation does not establish general semantic independence.

Test cases were inspected during development, without independent new-source confirmation. Costs are authored benchmark units; checks are idealized deterministic diagnostics. Each world has one saved response per evaluated model, so repeated-call variability is not measured. Finite-attacker recovery is empirical, not a formal privacy guarantee.

Citation

If you use TUPLE-Bench in your research, please cite:

@misc{yang2026tuplebench,
  title = {{TUPLE-Bench}: Evaluating and Mitigating Privacy-Utility
           Trade-offs in Typed Decision Models},
  author = {Yang, Yiqu and Zheng, Qiaoyuan and Tian, Yang
            and Fan, Yu and Ash, Elliott},
  year = {2026},
  eprint = {2610.23668},
  archivePrefix = {arXiv},
  primaryClass = {cs.CR},
  url = {https://arxiv.org/abs/2610.23668}
}

Download the BibTeX file.