PrivacyBench-Plus / README.md
adamrida's picture
PrivacyBench+ 1.0
4c52bbf
|
Raw History Blame Contribute Delete
7.35 kB
metadata
license: cc-by-4.0
language:
  - en
task_categories:
  - token-classification
task_ids:
  - named-entity-recognition
tags:
  - pii
  - privacy
  - redaction
  - anonymization
  - synthetic
pretty_name: PrivacyBench+
size_categories:
  - 10K<n<100K
source_datasets:
  - TonicAI/Privacy-Bench
configs:
  - config_name: messages
    default: true
    data_files:
      - data/human_annotations/daniel_lincoln_financial_group.jsonl
      - data/human_annotations/isabella_chipotle.jsonl
      - data/human_annotations/megan_donovan_eli_lilly.jsonl
      - data/human_annotations/morgan_morgan_stanley.jsonl
      - data/human_annotations/nora_caterpillar.jsonl
      - data/human_annotations/renee_tyson_foods.jsonl
  - config_name: documents
    data_files: data/human_annotations/*_documents.jsonl

PrivacyBench+

PrivacyBench+ is a benchmark for detecting personal and organization identifiers in workplace text. It contains 24,251 synthetic chat messages, emails and document pages written for fictional employees of 21 companies, with character-level span labels for ten entity types and a separate track for indirect identifiers.

It is a modified copy of PrivacyBench v2.5 by TonicAI (CC BY 4.0). The text is unchanged. The labels were revised against a written labelling policy, and indirect identifiers were added.

  • Maintained by: Tracer
  • Language: English
  • License: CC BY 4.0
PrivacyBench v2.5 PrivacyBench+
Records 24,251 24,251
Labels, human-annotated set (5,854 records, 6 companies) 25,855 26,212
Labels, native set (18,397 records, 21 companies) 130,370 144,192
Spans marked unresolved 0 1,861
Indirect identifiers 0 1,365

Dataset structure

Set Files Content
Human-annotated messages (config messages) data/human_annotations/<company>.jsonl 5,198 chat messages, 5 entity types
Human-annotated documents (config documents) data/human_annotations/<company>_documents.jsonl 656 document pages, 10 entity types
Native data/ground_truth/<company>/spans.jsonl 18,397 chat, email and document records for 21 companies, 10 entity types
from datasets import load_dataset

messages = load_dataset("Tracer-ML/PrivacyBench-Plus", "messages", split="train")
documents = load_dataset("Tracer-ML/PrivacyBench-Plus", "documents", split="train")

The native files are JSON Lines with metadata that varies by source format, and are read directly.

data/extensions/indirect_identifiers.jsonl lists the indirect identifiers for all three sets.

Fields

Field Description
text The message or page text
meta Source metadata (sender, recipients and time for messages; file name, format and page for documents)
ground_truth_spans List of {start, end, label, text} with character offsets into text
_tracer_review.unknown_targets Present on some records: spans whose status is unresolved

Entity types

Set Types
Messages NAME_GIVEN, NAME_FAMILY, ORGANIZATION, EMAIL_ADDRESS, USERNAME
Documents and native The five above, plus LOCATION_ADDRESS, ACCOUNT_NUMBER, PHONE_NUMBER, EMPLOYEE_ID, URL

Indirect identifiers are products, brands, programmes, tickers and sites that point to a single company, and standalone places. They are described in SCHEMA-EXTENSIONS.md.

Splits

The dataset has no fixed splits. The recommended evaluation set is the human-annotated records of two companies, megan_donovan_eli_lilly and nora_caterpillar (1,962 records). The native set has broader coverage and is suited to measuring recall.

Evaluation

python score.py data predictions.jsonl

predictions.jsonl has one line per record:

{"id": "<file name>:<line number>", "spans": [{"start": 0, "end": 5}]}

The scorer uses only the Python standard library and evaluates the two recommended companies by default. It reports two scores, both computed on masked characters with entity type ignored:

  • PII detection: precision, recall and F1 against the labels. Records with unresolved spans are excluded, and indirect identifiers are neutral.
  • Anonymization grade: labels and indirect identifiers are both required, on every record. Also reports the share of indirect identifiers hidden and the share of records with a required span left visible.

Results

PII detection on the recommended evaluation set. Full results are in SCORECARD.md.

System Precision Recall F1
GPT-6 Astra (prompted) 99.4 96.9 98.1
PIIX-560m 96.5 97.2 96.9
Fastino GLiNER2.5-base 89.3 83.4 86.2
Perplexity PII-Tracer 86.1 72.9 79.0
OpenAI Privacy Filter 87.6 64.1 74.1
Microsoft Presidio 52.6 83.7 64.6

PIIX-560m is developed by Tracer.

Dataset creation

Source data

The text comes from PrivacyBench v2.5 and is synthetic. The people are fictional and the companies are real names used in fictional scenarios. Record order and text are identical to the source, so labels can be joined to the original files by file name and line number. The original documents (emails, PDF, Word and spreadsheet files) and the generation tasks are available in the source dataset.

Annotation

Labels were revised under the labelling policy. Candidate corrections were proposed by automated audits and deterministic checks, then accepted or declined in review.

Review decisions Count
Made by a human reviewer 13,054
Applied automatically from the policy and the human reviewer's precedents, not individually checked by a person 15,494

Every decision is recorded with its reason in review/decisions.jsonl.gz. The main changes are summarised in CHANGELOG.md.

Limitations

  • The data is synthetic and English only.
  • The revised labels have not been independently re-annotated. Many corrections were applied in groups of identical changes.
  • 1,861 spans are marked unresolved and are excluded from scoring.
  • Indirect identifiers are not exhaustive. Job titles, dates and rare events are not covered.
  • Labels on the native set are less complete than on the human-annotated set. Use it for recall.
  • A score on this benchmark does not establish that a system anonymizes real data to a legal standard.

License and attribution

PrivacyBench+ is released under CC BY 4.0.

It is a modified copy of PrivacyBench v2.5 by TonicAI, used under CC BY 4.0. Modifications by Tracer: revised and added labels, spans marked unresolved, and an added file of indirect identifiers. PrivacyBench+ is not affiliated with or endorsed by TonicAI.

Citation

@misc{privacybenchplus2026,
  title  = {PrivacyBench+},
  author = {Tracer},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Tracer-ML/PrivacyBench-Plus}
}

Please also cite the source dataset, TonicAI PrivacyBench.