erchiw/DPriv-Bench

DPrivBench: Benchmarking LLMs’ Reasoning for Differential Privacy

DPrivBench is a benchmark for evaluating whether language models can correctly reason about and verify claimed differential privacy (DP) guarantees from natural-language/LaTeX-format problem statements.

This release contains evaluation data from seven benchmark configs, along with one auxiliary function bank:

• Category 1: 6 fundamental mechanism tracks, each with 98 questions.
• Category 2: 125 more advanced algorithm-level DP questions derived from literatures.

The data files are stored as <config_name>/test-*.parquet.

Example Usage

from datasets import load_dataset

repo = "erchiw/DPrivBench"
# Category 1
ds = load_dataset(repo, "cate_1_Laplace_pureDP", split="test")

# Category 2
ds = load_dataset(repo, "cate_2", split="test")

Evaluation code can be found in the GitHub Repository

Dataset Structure

Config	Description
cate_1_Laplace_pureDP	Laplace mechanism under pure DP
cate_1_Gaussian_GDP	Gaussian mechanism under GDP
cate_1_Gaussian_zCDP	Gaussian mechanism under zCDP
cate_1_ExpoMech_pureDP	Exponential mechanism under pure DP
cate_1_LaplaceRNM_pureDP	Report Noisy Max with Laplace noise
cate_1_PF_pureDP	Permute-and-Flip under pure DP
cate_1_function_bank	Function bank for category 1 questions
cate_2	algorithm-level DP questions

Task Format

Each example is a yes/no verification problem asking whether a mechanism or algorithm satisfies a claimed privacy guarantee under the stated assumptions.

Label formats

For all benchmark configs, the label field is an integer:

• 1 = yes
• 0 = no

Data Fields

Category 1 (cate_1_*)

Field	Description
question_id	Unique identifier for the question instance.
question	Text of the yes/no verification question.
label	Ground-truth label: 1 for yes, 0 for no.
function_id	Identifier of the query function in the function bank.
function	Mathematical definition of the function used in the question.
function_sens	L1 sensitivity of the function under a replace-one neighboring relation, assuming inputs in [0,1].

Category 2 (cate_2)

Field	Description
question_id	Unique identifier for the question instance.
question_tex	Full question statement in LaTeX format.
label	Ground-truth label: 1 for yes, 0 for no.
citation	Bibliographic citation(s) for the relevant paper(s). Multiple entries may be separated by ;.
negative_mode	Construction type of the example. "atom" denotes a base positive/negative question; other values indicate how a negative or counterexample-style question was derived.
pdf_link	URL or pointer to the referenced source document(s).
publish_year	Publication year of the primary references.
related_question	question_id of the related base question, when applicable. Missing values may appear as NaN.
section_number	Section or location in the source where the relevant claim appears.
subject	Coarse-grained subject area.
topic	Fine-grained topic label within the subject.
comments	Short proof sketch or rationale explaining the correctness of the label.

Please refer to the (rendered pdf) for a reader-friendly version.

Notes

• This release is test-only and is intended for evaluation rather than training.
• The cate_1_function_bank config is auxiliary and contains the function bank used to construct the Category 1 questions. For Category 1, we assume input data in $[0,1]$ and adopt the replace-one neighboring relation.
• In Category 2, the neighboring relation and input data range are specified case-by-case in each question.

Changelog

2026-06-12 — cate_2 corrections (questions 46, 47, 76)

Labels unchanged.

question_id	Field	Change
46	question_tex	Corrected added Gaussian noise variance from $\mathcal{N}(0, 2\sigma^2 C^2 I_d)$ to $\mathcal{N}(0, 4\sigma^2 C^2 I_d)$.
47	question_tex	Corrected added Gaussian noise variance from $\mathcal{N}(0, 2\sigma^2 I_d)$ to $\mathcal{N}(0, 4\sigma^2 I_d)$.
47	comments	Updated $L_2$ sensitivity of the clipped gradient sum from $\sqrt{2}$ to $2$.
76	comments	Clarified that the query can induce sensitivity larger than $C$ (replacing the previous $2C$ wording) under the add/remove neighboring relation.

2026-04 — Initial release

First public release of DPrivBench on Hugging Face, including:

• Category 1: 6 mechanism tracks (cate_1_*), 98 questions each, plus the cate_1_function_bank auxiliary config.
• Category 2: 125 algorithm-level DP questions (cate_2).

Citation

If you use this dataset, please cite the DPrivBench paper.

@misc{dprivbenchauthors,
  title={DPrivBench: Benchmarking LLMs' Reasoning for Differential Privacy},
  author={Erchi Wang and Pengrun Huang and Eli Chien and Om Thakkar and Kamalika Chaudhuri and Yu-Xiang Wang and Ruihan Wu},
  year={2026},
  eprint={2604.15851},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2604.15851}, 
}