jayluxferro/llm-redactor-leak-benchmark

LLM-Redactor Leak Benchmark

A benchmark of 1,300 synthetic prompts with 4,014 ground-truth annotations spanning four workload classes, designed to evaluate privacy-preserving techniques for outbound LLM requests.

Released alongside the paper:

LLM-Redactor: An Empirical Evaluation of Eight Techniques for Privacy-Preserving LLM Requests

Justice Owusu Agyemang, Jerry John Kponyo, Elliot Amponsah, Godfred Manu Addo Boakye, Kwame Opuni-Boachie Obour Agyekum

arXiv:2604.12064

Workload classes

Config	Samples	Description
wl1_pii	500	Names, emails, phone numbers, addresses, SSNs, employee IDs
wl2_secrets	300	API keys, AWS credentials, passwords, hostnames in configs/code
wl3_implicit	200	Indirect references that identify people or organisations
wl4_code	300	Internal function names, database schemas, project names

Usage

from datasets import load_dataset

# Load everything
ds = load_dataset("jayluxferro/llm-redactor-leak-benchmark")

# Load a single workload
pii = load_dataset("jayluxferro/llm-redactor-leak-benchmark", "wl1_pii")

Schema

Each sample has the following fields:

Field	Type	Description
id	string	Unique identifier (e.g. wl1_0042)
text	string	The input prompt to evaluate
workload	string	Workload class (wl1_pii, wl2_secrets, wl3_implicit, wl4_code)
annotations	list[object]	Ground-truth sensitive spans
annotations[].start	int	Start character offset
annotations[].end	int	End character offset
annotations[].kind	string	Sensitivity type (e.g. person, email, api_key, implicit)
annotations[].text	string	The verbatim sensitive span

Annotation kinds

WL1 (PII): person, email, phone, address, ssn, employee_id, org_name

WL2 (Secrets): aws_access_key, aws_secret_key, api_key, password, hostname

WL3 (Implicit): implicit, org_name

WL4 (Code): project_name, org_name, internal_function, database_name, table_name, api_key, hostname

Citation

@article{agyemang2026llmredactor,
  title={LLM-Redactor: An Empirical Evaluation of Eight Techniques for Privacy-Preserving LLM Requests},
  author={Agyemang, Justice Owusu and Kponyo, Jerry John and Amponsah, Elliot and Boakye, Godfred Manu Addo and Agyekum, Kwame Opuni-Boachie Obour},
  year={2026},
  url={https://arxiv.org/abs/2604.12064}
}

License

MIT