Datasets:
Tasks:
Token Classification
Modalities:
Text
Formats:
json
Sub-tasks:
named-entity-recognition
Languages:
English
Size:
1K - 10K
License:
Commit ·
4c52bbf
0
Parent(s):
PrivacyBench+ 1.0
Browse files- .gitattributes +60 -0
- CHANGELOG.md +41 -0
- MANIFEST.json +44 -0
- NOTICE.md +14 -0
- POLICY.md +47 -0
- README.md +186 -0
- SCHEMA-EXTENSIONS.md +50 -0
- SCORECARD.md +59 -0
- data/extensions/indirect_identifiers.jsonl +0 -0
- data/ground_truth/aaron_pfizer/spans.jsonl +0 -0
- data/ground_truth/camille_nike/spans.jsonl +0 -0
- data/ground_truth/daniel_lincoln_financial_group/spans.jsonl +0 -0
- data/ground_truth/devon_marriott_international/spans.jsonl +0 -0
- data/ground_truth/elena_stripe/spans.jsonl +0 -0
- data/ground_truth/grace_deere_company/spans.jsonl +0 -0
- data/ground_truth/hannah_salesforce/spans.jsonl +0 -0
- data/ground_truth/isabella_chipotle/spans.jsonl +0 -0
- data/ground_truth/malik_spotify/spans.jsonl +0 -0
- data/ground_truth/marcus_boeing/spans.jsonl +0 -0
- data/ground_truth/megan_donovan_eli_lilly/spans.jsonl +0 -0
- data/ground_truth/morgan_morgan_stanley/spans.jsonl +0 -0
- data/ground_truth/nora_caterpillar/spans.jsonl +0 -0
- data/ground_truth/priya_charles_schwab/spans.jsonl +0 -0
- data/ground_truth/ray_raymond_james/spans.jsonl +0 -0
- data/ground_truth/renee_tyson_foods/spans.jsonl +0 -0
- data/ground_truth/sofia_netflix/spans.jsonl +0 -0
- data/ground_truth/tanya_kroger/spans.jsonl +0 -0
- data/ground_truth/victor_delta_air_lines/spans.jsonl +0 -0
- data/ground_truth/wells_wells_fargo/spans.jsonl +0 -0
- data/ground_truth/yusuf_procter_gamble/spans.jsonl +0 -0
- data/human_annotations/daniel_lincoln_financial_group.jsonl +0 -0
- data/human_annotations/daniel_lincoln_financial_group_documents.jsonl +0 -0
- data/human_annotations/isabella_chipotle.jsonl +0 -0
- data/human_annotations/isabella_chipotle_documents.jsonl +0 -0
- data/human_annotations/megan_donovan_eli_lilly.jsonl +0 -0
- data/human_annotations/megan_donovan_eli_lilly_documents.jsonl +0 -0
- data/human_annotations/morgan_morgan_stanley.jsonl +0 -0
- data/human_annotations/morgan_morgan_stanley_documents.jsonl +0 -0
- data/human_annotations/nora_caterpillar.jsonl +0 -0
- data/human_annotations/nora_caterpillar_documents.jsonl +0 -0
- data/human_annotations/renee_tyson_foods.jsonl +0 -0
- data/human_annotations/renee_tyson_foods_documents.jsonl +0 -0
- review/decisions.jsonl.gz +3 -0
- score.py +71 -0
.gitattributes
ADDED
|
@@ -0,0 +1,60 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
*.7z filter=lfs diff=lfs merge=lfs -text
|
| 2 |
+
*.arrow filter=lfs diff=lfs merge=lfs -text
|
| 3 |
+
*.avro filter=lfs diff=lfs merge=lfs -text
|
| 4 |
+
*.bin filter=lfs diff=lfs merge=lfs -text
|
| 5 |
+
*.bz2 filter=lfs diff=lfs merge=lfs -text
|
| 6 |
+
*.ckpt filter=lfs diff=lfs merge=lfs -text
|
| 7 |
+
*.ftz filter=lfs diff=lfs merge=lfs -text
|
| 8 |
+
*.gz filter=lfs diff=lfs merge=lfs -text
|
| 9 |
+
*.h5 filter=lfs diff=lfs merge=lfs -text
|
| 10 |
+
*.joblib filter=lfs diff=lfs merge=lfs -text
|
| 11 |
+
*.lfs.* filter=lfs diff=lfs merge=lfs -text
|
| 12 |
+
*.lz4 filter=lfs diff=lfs merge=lfs -text
|
| 13 |
+
*.mds filter=lfs diff=lfs merge=lfs -text
|
| 14 |
+
*.mlmodel filter=lfs diff=lfs merge=lfs -text
|
| 15 |
+
*.model filter=lfs diff=lfs merge=lfs -text
|
| 16 |
+
*.msgpack filter=lfs diff=lfs merge=lfs -text
|
| 17 |
+
*.npy filter=lfs diff=lfs merge=lfs -text
|
| 18 |
+
*.npz filter=lfs diff=lfs merge=lfs -text
|
| 19 |
+
*.onnx filter=lfs diff=lfs merge=lfs -text
|
| 20 |
+
*.ot filter=lfs diff=lfs merge=lfs -text
|
| 21 |
+
*.parquet filter=lfs diff=lfs merge=lfs -text
|
| 22 |
+
*.pb filter=lfs diff=lfs merge=lfs -text
|
| 23 |
+
*.pickle filter=lfs diff=lfs merge=lfs -text
|
| 24 |
+
*.pkl filter=lfs diff=lfs merge=lfs -text
|
| 25 |
+
*.pt filter=lfs diff=lfs merge=lfs -text
|
| 26 |
+
*.pth filter=lfs diff=lfs merge=lfs -text
|
| 27 |
+
*.rar filter=lfs diff=lfs merge=lfs -text
|
| 28 |
+
*.safetensors filter=lfs diff=lfs merge=lfs -text
|
| 29 |
+
saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
| 30 |
+
*.tar.* filter=lfs diff=lfs merge=lfs -text
|
| 31 |
+
*.tar filter=lfs diff=lfs merge=lfs -text
|
| 32 |
+
*.tflite filter=lfs diff=lfs merge=lfs -text
|
| 33 |
+
*.tgz filter=lfs diff=lfs merge=lfs -text
|
| 34 |
+
*.wasm filter=lfs diff=lfs merge=lfs -text
|
| 35 |
+
*.xz filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 37 |
+
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 38 |
+
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 39 |
+
# Audio files - uncompressed
|
| 40 |
+
*.pcm filter=lfs diff=lfs merge=lfs -text
|
| 41 |
+
*.sam filter=lfs diff=lfs merge=lfs -text
|
| 42 |
+
*.raw filter=lfs diff=lfs merge=lfs -text
|
| 43 |
+
# Audio files - compressed
|
| 44 |
+
*.aac filter=lfs diff=lfs merge=lfs -text
|
| 45 |
+
*.flac filter=lfs diff=lfs merge=lfs -text
|
| 46 |
+
*.mp3 filter=lfs diff=lfs merge=lfs -text
|
| 47 |
+
*.ogg filter=lfs diff=lfs merge=lfs -text
|
| 48 |
+
*.wav filter=lfs diff=lfs merge=lfs -text
|
| 49 |
+
# Image files - uncompressed
|
| 50 |
+
*.bmp filter=lfs diff=lfs merge=lfs -text
|
| 51 |
+
*.gif filter=lfs diff=lfs merge=lfs -text
|
| 52 |
+
*.png filter=lfs diff=lfs merge=lfs -text
|
| 53 |
+
*.tiff filter=lfs diff=lfs merge=lfs -text
|
| 54 |
+
# Image files - compressed
|
| 55 |
+
*.jpg filter=lfs diff=lfs merge=lfs -text
|
| 56 |
+
*.jpeg filter=lfs diff=lfs merge=lfs -text
|
| 57 |
+
*.webp filter=lfs diff=lfs merge=lfs -text
|
| 58 |
+
# Video files - compressed
|
| 59 |
+
*.mp4 filter=lfs diff=lfs merge=lfs -text
|
| 60 |
+
*.webm filter=lfs diff=lfs merge=lfs -text
|
CHANGELOG.md
ADDED
|
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Changelog
|
| 2 |
+
|
| 3 |
+
## 1.0 (November 2026)
|
| 4 |
+
|
| 5 |
+
First release. Changes relative to PrivacyBench v2.5:
|
| 6 |
+
|
| 7 |
+
| | PrivacyBench v2.5 | PrivacyBench+ 1.0 |
|
| 8 |
+
|---|---|---|
|
| 9 |
+
| Records | 24,251 | 24,251 |
|
| 10 |
+
| Labels, human-annotated set | 25,855 | 26,212 |
|
| 11 |
+
| Labels, native set | 130,370 | 144,192 |
|
| 12 |
+
| Spans marked unresolved | 0 | 1,861 |
|
| 13 |
+
| Indirect identifiers | 0 | 1,365 |
|
| 14 |
+
|
| 15 |
+
### Added
|
| 16 |
+
|
| 17 |
+
- Missing names, including initials and monograms.
|
| 18 |
+
- Missing organizations, email addresses and account numbers.
|
| 19 |
+
- Indirect identifiers, in `data/extensions/indirect_identifiers.jsonl`.
|
| 20 |
+
- A scorer, `score.py`, with a second score that requires indirect identifiers.
|
| 21 |
+
|
| 22 |
+
### Changed
|
| 23 |
+
|
| 24 |
+
- Organization spans include the legal suffix and exclude attached departments and punctuation.
|
| 25 |
+
- Street-only addresses are extended to the full postal address.
|
| 26 |
+
- Labels with the wrong type are retyped: person and organization homonyms, and given and family names.
|
| 27 |
+
|
| 28 |
+
### Removed
|
| 29 |
+
|
| 30 |
+
- Organization and name labels on departments, internal tools, public bodies, products and chat slang. Products,
|
| 31 |
+
brands, programmes, tickers and named sites that point to one company are recorded as indirect identifiers.
|
| 32 |
+
|
| 33 |
+
### Review record
|
| 34 |
+
|
| 35 |
+
28,548 review decisions: 13,054 made by a human reviewer and 15,494 applied automatically from the labelling policy
|
| 36 |
+
and the human reviewer's precedents. Of the automatic decisions, 13,200 applied a correction, 1,964 declined a
|
| 37 |
+
proposal and 330 left the case unresolved. All decisions are in `review/decisions.jsonl.gz`.
|
| 38 |
+
|
| 39 |
+
### Source version
|
| 40 |
+
|
| 41 |
+
Based on PrivacyBench v2.5 at revision `a9f3e392`. Later changes to the source dataset are not included.
|
MANIFEST.json
ADDED
|
@@ -0,0 +1,44 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"CHANGELOG.md": "98e3d691e2f70a3efe447d2de7537a8811b034898b853ad9908055f4825611f3",
|
| 3 |
+
"NOTICE.md": "5954fa2ef2c8042c1ad3309af0c3eb6bf41c47f5cde31cd0e63f423e86ab25dd",
|
| 4 |
+
"POLICY.md": "cf2b9beba809a91ad273a9814800d9cb6abe8af385ed766782b012cde6897987",
|
| 5 |
+
"README.md": "bd96c0f992e3076d4525905cc422434442fa9fdc23cbd926092464f19b59f476",
|
| 6 |
+
"SCHEMA-EXTENSIONS.md": "ec3dbf4bc682e10ef3dd24f269d9a9524606d9203116ba336dc31fd610611d2a",
|
| 7 |
+
"SCORECARD.md": "3a3b7f75384951c8e2b8913636cbff24fdffd854bc4859d8b145317c49e7815f",
|
| 8 |
+
"data/extensions/indirect_identifiers.jsonl": "0a6fb22fd76da47ae84124250b8ac50382ac31f8d86112f22110c7fdcdba3927",
|
| 9 |
+
"data/ground_truth/aaron_pfizer/spans.jsonl": "956d98d10356e8fa48249da1858cd21d7a79f24f37bb359341e728760155764b",
|
| 10 |
+
"data/ground_truth/camille_nike/spans.jsonl": "edaf50c8c2ddfe810b19a91bd27462d8065efd02ad5079231524e94171388e92",
|
| 11 |
+
"data/ground_truth/daniel_lincoln_financial_group/spans.jsonl": "5857a53f31fa504a80ff58ede1c867ca84a037e7387de517dcc10e61e56e501c",
|
| 12 |
+
"data/ground_truth/devon_marriott_international/spans.jsonl": "fd5032f4227a754de184322a291eb83f05fd18ce23c549fd2a9970bcee8c1cf0",
|
| 13 |
+
"data/ground_truth/elena_stripe/spans.jsonl": "909b552279a71aee297763211e6d196088345d0d6ada392454d6d0953cbad0c8",
|
| 14 |
+
"data/ground_truth/grace_deere_company/spans.jsonl": "4f1a0e3e4b2ffcaf19cd93c095d0e994aa3d27b39ff013d413aabc3b29cb25ae",
|
| 15 |
+
"data/ground_truth/hannah_salesforce/spans.jsonl": "6892b7b3d390c48700afdb0d9b637d945c65a5773637ff0263d4194138b077af",
|
| 16 |
+
"data/ground_truth/isabella_chipotle/spans.jsonl": "addc16442d06df54550bab5d6f9d307918151bb69f4467bdef20b9616018bcf5",
|
| 17 |
+
"data/ground_truth/malik_spotify/spans.jsonl": "d17f0059fbc02eb9c9b5ddde2e0b0734cea23914b4a052af3badfaf51283299f",
|
| 18 |
+
"data/ground_truth/marcus_boeing/spans.jsonl": "f4c2608c812fe7d1fc9935ac2d951803baa6eb3cb911f1f082ca5de3e308acfd",
|
| 19 |
+
"data/ground_truth/megan_donovan_eli_lilly/spans.jsonl": "6fd5019d8f8be8597a9632df990fc9868e0b162ef727abbdfbe6cc5f4c20beb6",
|
| 20 |
+
"data/ground_truth/morgan_morgan_stanley/spans.jsonl": "514cbf54b4bd928a8e09031ff7b666f6f019930757cd937a51072d68210ad36d",
|
| 21 |
+
"data/ground_truth/nora_caterpillar/spans.jsonl": "a75a2350b4c75ed8aa8bcd85f05770c7f9c2cc3b80a8346a84d0c3df3c3c55a4",
|
| 22 |
+
"data/ground_truth/priya_charles_schwab/spans.jsonl": "2bb2d3a01c7cbb355af7f5b1a65b1b8dfe54aa1b608277f5f5d59b7a96e88653",
|
| 23 |
+
"data/ground_truth/ray_raymond_james/spans.jsonl": "b4eb1082865e61d0796022047e554459dfb22429163a4bd26c81f4e4c41ba8c9",
|
| 24 |
+
"data/ground_truth/renee_tyson_foods/spans.jsonl": "c00ec718547971c9c7d63020e9dcad366da11b407f2bbc8cb4617614ff86081f",
|
| 25 |
+
"data/ground_truth/sofia_netflix/spans.jsonl": "6b80c25667b34a28cb54c2c7e74dc067595ae663c7e5e2a565cbf662b6e11ae2",
|
| 26 |
+
"data/ground_truth/tanya_kroger/spans.jsonl": "70dc52408aa89fea7cc266ae4da9fc725d1737e046f690b3a984fe91bc6e3bab",
|
| 27 |
+
"data/ground_truth/victor_delta_air_lines/spans.jsonl": "32a90896c3c0c7979ac27cf97b5ea9ae8e6754da0b4030c33d9f991fb73fab22",
|
| 28 |
+
"data/ground_truth/wells_wells_fargo/spans.jsonl": "f9387ade8fe3135371332fb01f5e3da49ca35c4dab3d9449a0255217903f2b4c",
|
| 29 |
+
"data/ground_truth/yusuf_procter_gamble/spans.jsonl": "afecd65e5707a4ecf542e2adde36c69df6168937456f69242355797e8886b0a0",
|
| 30 |
+
"data/human_annotations/daniel_lincoln_financial_group.jsonl": "296889c9a24c595ca7b746c965bfdf9ea4907da7f5567c5fc438cdf187adb616",
|
| 31 |
+
"data/human_annotations/daniel_lincoln_financial_group_documents.jsonl": "e59c09d8e1da77bd5760b3abfabc2e484596b44ddca701c25b6737bd20956317",
|
| 32 |
+
"data/human_annotations/isabella_chipotle.jsonl": "408bd1fffb717cb688e0cc4311cd4f794895dbbfbbbf569884bcb0492802d1ce",
|
| 33 |
+
"data/human_annotations/isabella_chipotle_documents.jsonl": "3922206fd8b83ab30ce9eab2686e35e065d960f084c09e2c45a674c6ef0af5b7",
|
| 34 |
+
"data/human_annotations/megan_donovan_eli_lilly.jsonl": "a2ab019a54a11cfefd56037d28b78104a3e32c9fd22f3c1d7b66a2aaebe87ee8",
|
| 35 |
+
"data/human_annotations/megan_donovan_eli_lilly_documents.jsonl": "bc439e438d55118620d531a589ffbea5cbbf9f62414d03229a9c96209c96604d",
|
| 36 |
+
"data/human_annotations/morgan_morgan_stanley.jsonl": "ca2f91a104a348ddaad7f455be772b3755aeb649a4477fcbd877478b1580de37",
|
| 37 |
+
"data/human_annotations/morgan_morgan_stanley_documents.jsonl": "6a88089caff770293f65d42b4171501701004eaa402b44b2f3129a24fd26e1ed",
|
| 38 |
+
"data/human_annotations/nora_caterpillar.jsonl": "f59912ba96fd09ec1993cdd38400fd45fabc79b1cb3eda6bcbe508411bcd5a9d",
|
| 39 |
+
"data/human_annotations/nora_caterpillar_documents.jsonl": "ef3c1de7b00382a9f063facb2b8ba970bbf43dfda2a40b86bf099139935c424d",
|
| 40 |
+
"data/human_annotations/renee_tyson_foods.jsonl": "327859a29e758790ee834662d73aaf66e4d604a4de1ec5f86127a9afb6bdfe31",
|
| 41 |
+
"data/human_annotations/renee_tyson_foods_documents.jsonl": "428a327327a209136046e2fd1256432b8c856d1f87a56cb6e67525d46030c1d0",
|
| 42 |
+
"review/decisions.jsonl.gz": "d18354857796387a26ac016fe9d50e2e0685baf200c31c99e30d154d5616bf0e",
|
| 43 |
+
"score.py": "29b6f85e90e0d1ed902e641852712e6882c5585228f5335db0c920dfe0712235"
|
| 44 |
+
}
|
NOTICE.md
ADDED
|
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Notice
|
| 2 |
+
|
| 3 |
+
PrivacyBench+
|
| 4 |
+
Copyright 2026 Tracer. Licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).
|
| 5 |
+
https://creativecommons.org/licenses/by/4.0/
|
| 6 |
+
|
| 7 |
+
This dataset is a modified copy of PrivacyBench v2.5 by TonicAI, licensed under CC BY 4.0.
|
| 8 |
+
https://huggingface.co/datasets/TonicAI/Privacy-Bench
|
| 9 |
+
|
| 10 |
+
Modifications by Tracer: revised and added labels, spans marked unresolved, and an added file of indirect
|
| 11 |
+
identifiers. The text of the records is unchanged. The changes are summarised in `CHANGELOG.md` and listed in
|
| 12 |
+
`review/decisions.jsonl.gz`.
|
| 13 |
+
|
| 14 |
+
PrivacyBench+ is not affiliated with or endorsed by TonicAI.
|
POLICY.md
ADDED
|
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Labelling policy
|
| 2 |
+
|
| 3 |
+
This policy defines which spans are labelled in PrivacyBench+. Labels are character spans over the unchanged source
|
| 4 |
+
text.
|
| 5 |
+
|
| 6 |
+
## Principles
|
| 7 |
+
|
| 8 |
+
1. Direct identifiers of a person are always labelled.
|
| 9 |
+
2. An organization is labelled when it is the employer, client, customer or counterparty of the people in the text.
|
| 10 |
+
3. Names that do not identify a specific company or person are not labelled.
|
| 11 |
+
4. Names that reveal a company without naming it are recorded as indirect identifiers, outside the main labels.
|
| 12 |
+
5. When the text does not settle a case, the span is marked unresolved and is excluded from scoring.
|
| 13 |
+
|
| 14 |
+
## Rules by type
|
| 15 |
+
|
| 16 |
+
| Type | Labelled | Not labelled |
|
| 17 |
+
|---|---|---|
|
| 18 |
+
| `NAME_GIVEN`, `NAME_FAMILY` | Given and family names of people, including initials and monograms that stand for a name | Chat slang and common words that match a name |
|
| 19 |
+
| `ORGANIZATION` | Employers, clients, customers and counterparties. The span includes the legal suffix and excludes attached department names and punctuation | Internal tools (for example Slack, Zoom, SAP), public bodies (for example SEC, IRS), generic departments (for example Human Resources), products and brands |
|
| 20 |
+
| `EMAIL_ADDRESS`, `USERNAME`, `URL` | Always | |
|
| 21 |
+
| `PHONE_NUMBER`, `ACCOUNT_NUMBER`, `EMPLOYEE_ID` | Always | |
|
| 22 |
+
| `LOCATION_ADDRESS` | Postal addresses, with every component present in the text: street, city, state or region, postal code and country | A state, region or country mentioned outside an address |
|
| 23 |
+
|
| 24 |
+
Further rules:
|
| 25 |
+
|
| 26 |
+
- A tool name used as an employer or client is labelled as an organization. For example, Salesforce is labelled when
|
| 27 |
+
it is the employer or a client, and is not labelled in "log it in Salesforce".
|
| 28 |
+
- The tool exception applies to the bare name only. An account, username, email address or URL that contains the
|
| 29 |
+
name is still labelled.
|
| 30 |
+
- The type of a label follows its meaning in context. A person named Jordan is a name, not a location.
|
| 31 |
+
|
| 32 |
+
## Indirect identifiers
|
| 33 |
+
|
| 34 |
+
Products, brands, programmes, tickers and named sites that belong to one company, and standalone cities, sites and
|
| 35 |
+
venues, are recorded in `data/extensions/indirect_identifiers.jsonl`. They are not part of `ground_truth_spans`. See
|
| 36 |
+
[SCHEMA-EXTENSIONS.md](SCHEMA-EXTENSIONS.md).
|
| 37 |
+
|
| 38 |
+
## Unresolved spans
|
| 39 |
+
|
| 40 |
+
A span is marked unresolved when neither this policy nor the context decides it, for example a bare name in a
|
| 41 |
+
signature that could be a tool or a client. Unresolved spans are listed under `_tracer_review.unknown_targets`. They
|
| 42 |
+
are never treated as negatives: masking them is neither rewarded nor penalised.
|
| 43 |
+
|
| 44 |
+
## Label scope
|
| 45 |
+
|
| 46 |
+
Chat messages in the human-annotated set carry five types (names, organization, email address, username). Documents
|
| 47 |
+
and the native set carry all ten.
|
README.md
ADDED
|
@@ -0,0 +1,186 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-4.0
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
task_categories:
|
| 6 |
+
- token-classification
|
| 7 |
+
task_ids:
|
| 8 |
+
- named-entity-recognition
|
| 9 |
+
tags:
|
| 10 |
+
- pii
|
| 11 |
+
- privacy
|
| 12 |
+
- redaction
|
| 13 |
+
- anonymization
|
| 14 |
+
- synthetic
|
| 15 |
+
pretty_name: PrivacyBench+
|
| 16 |
+
size_categories:
|
| 17 |
+
- 10K<n<100K
|
| 18 |
+
source_datasets:
|
| 19 |
+
- TonicAI/Privacy-Bench
|
| 20 |
+
configs:
|
| 21 |
+
- config_name: messages
|
| 22 |
+
default: true
|
| 23 |
+
data_files:
|
| 24 |
+
- data/human_annotations/daniel_lincoln_financial_group.jsonl
|
| 25 |
+
- data/human_annotations/isabella_chipotle.jsonl
|
| 26 |
+
- data/human_annotations/megan_donovan_eli_lilly.jsonl
|
| 27 |
+
- data/human_annotations/morgan_morgan_stanley.jsonl
|
| 28 |
+
- data/human_annotations/nora_caterpillar.jsonl
|
| 29 |
+
- data/human_annotations/renee_tyson_foods.jsonl
|
| 30 |
+
- config_name: documents
|
| 31 |
+
data_files: "data/human_annotations/*_documents.jsonl"
|
| 32 |
+
---
|
| 33 |
+
|
| 34 |
+
# PrivacyBench+
|
| 35 |
+
|
| 36 |
+
PrivacyBench+ is a benchmark for detecting personal and organization identifiers in workplace text. It contains
|
| 37 |
+
24,251 synthetic chat messages, emails and document pages written for fictional employees of 21 companies, with
|
| 38 |
+
character-level span labels for ten entity types and a separate track for indirect identifiers.
|
| 39 |
+
|
| 40 |
+
It is a modified copy of [PrivacyBench v2.5](https://huggingface.co/datasets/TonicAI/Privacy-Bench) by TonicAI
|
| 41 |
+
(CC BY 4.0). The text is unchanged. The labels were revised against a written [labelling policy](POLICY.md), and
|
| 42 |
+
indirect identifiers were added.
|
| 43 |
+
|
| 44 |
+
- **Maintained by:** [Tracer](https://tracerml.ai)
|
| 45 |
+
- **Language:** English
|
| 46 |
+
- **License:** CC BY 4.0
|
| 47 |
+
|
| 48 |
+
| | PrivacyBench v2.5 | PrivacyBench+ |
|
| 49 |
+
|---|---|---|
|
| 50 |
+
| Records | 24,251 | 24,251 |
|
| 51 |
+
| Labels, human-annotated set (5,854 records, 6 companies) | 25,855 | 26,212 |
|
| 52 |
+
| Labels, native set (18,397 records, 21 companies) | 130,370 | 144,192 |
|
| 53 |
+
| Spans marked unresolved | 0 | 1,861 |
|
| 54 |
+
| Indirect identifiers | 0 | 1,365 |
|
| 55 |
+
|
| 56 |
+
## Dataset structure
|
| 57 |
+
|
| 58 |
+
| Set | Files | Content |
|
| 59 |
+
|---|---|---|
|
| 60 |
+
| Human-annotated messages (config `messages`) | `data/human_annotations/<company>.jsonl` | 5,198 chat messages, 5 entity types |
|
| 61 |
+
| Human-annotated documents (config `documents`) | `data/human_annotations/<company>_documents.jsonl` | 656 document pages, 10 entity types |
|
| 62 |
+
| Native | `data/ground_truth/<company>/spans.jsonl` | 18,397 chat, email and document records for 21 companies, 10 entity types |
|
| 63 |
+
|
| 64 |
+
```python
|
| 65 |
+
from datasets import load_dataset
|
| 66 |
+
|
| 67 |
+
messages = load_dataset("Tracer-ML/PrivacyBench-Plus", "messages", split="train")
|
| 68 |
+
documents = load_dataset("Tracer-ML/PrivacyBench-Plus", "documents", split="train")
|
| 69 |
+
```
|
| 70 |
+
|
| 71 |
+
The native files are JSON Lines with metadata that varies by source format, and are read directly.
|
| 72 |
+
|
| 73 |
+
`data/extensions/indirect_identifiers.jsonl` lists the indirect identifiers for all three sets.
|
| 74 |
+
|
| 75 |
+
### Fields
|
| 76 |
+
|
| 77 |
+
| Field | Description |
|
| 78 |
+
|---|---|
|
| 79 |
+
| `text` | The message or page text |
|
| 80 |
+
| `meta` | Source metadata (sender, recipients and time for messages; file name, format and page for documents) |
|
| 81 |
+
| `ground_truth_spans` | List of `{start, end, label, text}` with character offsets into `text` |
|
| 82 |
+
| `_tracer_review.unknown_targets` | Present on some records: spans whose status is unresolved |
|
| 83 |
+
|
| 84 |
+
### Entity types
|
| 85 |
+
|
| 86 |
+
| Set | Types |
|
| 87 |
+
|---|---|
|
| 88 |
+
| Messages | `NAME_GIVEN`, `NAME_FAMILY`, `ORGANIZATION`, `EMAIL_ADDRESS`, `USERNAME` |
|
| 89 |
+
| Documents and native | The five above, plus `LOCATION_ADDRESS`, `ACCOUNT_NUMBER`, `PHONE_NUMBER`, `EMPLOYEE_ID`, `URL` |
|
| 90 |
+
|
| 91 |
+
Indirect identifiers are products, brands, programmes, tickers and sites that point to a single company, and
|
| 92 |
+
standalone places. They are described in [SCHEMA-EXTENSIONS.md](SCHEMA-EXTENSIONS.md).
|
| 93 |
+
|
| 94 |
+
### Splits
|
| 95 |
+
|
| 96 |
+
The dataset has no fixed splits. The recommended evaluation set is the human-annotated records of two companies,
|
| 97 |
+
`megan_donovan_eli_lilly` and `nora_caterpillar` (1,962 records). The native set has broader coverage and is suited to
|
| 98 |
+
measuring recall.
|
| 99 |
+
|
| 100 |
+
## Evaluation
|
| 101 |
+
|
| 102 |
+
```bash
|
| 103 |
+
python score.py data predictions.jsonl
|
| 104 |
+
```
|
| 105 |
+
|
| 106 |
+
`predictions.jsonl` has one line per record:
|
| 107 |
+
|
| 108 |
+
```json
|
| 109 |
+
{"id": "<file name>:<line number>", "spans": [{"start": 0, "end": 5}]}
|
| 110 |
+
```
|
| 111 |
+
|
| 112 |
+
The scorer uses only the Python standard library and evaluates the two recommended companies by default. It reports
|
| 113 |
+
two scores, both computed on masked characters with entity type ignored:
|
| 114 |
+
|
| 115 |
+
- **PII detection:** precision, recall and F1 against the labels. Records with unresolved spans are excluded, and
|
| 116 |
+
indirect identifiers are neutral.
|
| 117 |
+
- **Anonymization grade:** labels and indirect identifiers are both required, on every record. Also reports the share
|
| 118 |
+
of indirect identifiers hidden and the share of records with a required span left visible.
|
| 119 |
+
|
| 120 |
+
## Results
|
| 121 |
+
|
| 122 |
+
PII detection on the recommended evaluation set. Full results are in [SCORECARD.md](SCORECARD.md).
|
| 123 |
+
|
| 124 |
+
| System | Precision | Recall | F1 |
|
| 125 |
+
|---|---|---|---|
|
| 126 |
+
| GPT-6 Astra (prompted) | 99.4 | 96.9 | 98.1 |
|
| 127 |
+
| [PIIX-560m](https://huggingface.co/Tracer-ML/PIIX-560m) | 96.5 | 97.2 | 96.9 |
|
| 128 |
+
| Fastino GLiNER2.5-base | 89.3 | 83.4 | 86.2 |
|
| 129 |
+
| Perplexity PII-Tracer | 86.1 | 72.9 | 79.0 |
|
| 130 |
+
| OpenAI Privacy Filter | 87.6 | 64.1 | 74.1 |
|
| 131 |
+
| Microsoft Presidio | 52.6 | 83.7 | 64.6 |
|
| 132 |
+
|
| 133 |
+
PIIX-560m is developed by Tracer.
|
| 134 |
+
|
| 135 |
+
## Dataset creation
|
| 136 |
+
|
| 137 |
+
### Source data
|
| 138 |
+
|
| 139 |
+
The text comes from PrivacyBench v2.5 and is synthetic. The people are fictional and the companies are real names
|
| 140 |
+
used in fictional scenarios. Record order and text are identical to the source, so labels can be joined to the
|
| 141 |
+
original files by file name and line number. The original documents (emails, PDF, Word and spreadsheet files) and the
|
| 142 |
+
generation tasks are available in the [source dataset](https://huggingface.co/datasets/TonicAI/Privacy-Bench).
|
| 143 |
+
|
| 144 |
+
### Annotation
|
| 145 |
+
|
| 146 |
+
Labels were revised under the [labelling policy](POLICY.md). Candidate corrections were proposed by automated audits
|
| 147 |
+
and deterministic checks, then accepted or declined in review.
|
| 148 |
+
|
| 149 |
+
| Review decisions | Count |
|
| 150 |
+
|---|---|
|
| 151 |
+
| Made by a human reviewer | 13,054 |
|
| 152 |
+
| Applied automatically from the policy and the human reviewer's precedents, not individually checked by a person | 15,494 |
|
| 153 |
+
|
| 154 |
+
Every decision is recorded with its reason in `review/decisions.jsonl.gz`. The main changes are summarised in
|
| 155 |
+
[CHANGELOG.md](CHANGELOG.md).
|
| 156 |
+
|
| 157 |
+
## Limitations
|
| 158 |
+
|
| 159 |
+
- The data is synthetic and English only.
|
| 160 |
+
- The revised labels have not been independently re-annotated. Many corrections were applied in groups of identical
|
| 161 |
+
changes.
|
| 162 |
+
- 1,861 spans are marked unresolved and are excluded from scoring.
|
| 163 |
+
- Indirect identifiers are not exhaustive. Job titles, dates and rare events are not covered.
|
| 164 |
+
- Labels on the native set are less complete than on the human-annotated set. Use it for recall.
|
| 165 |
+
- A score on this benchmark does not establish that a system anonymizes real data to a legal standard.
|
| 166 |
+
|
| 167 |
+
## License and attribution
|
| 168 |
+
|
| 169 |
+
PrivacyBench+ is released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
|
| 170 |
+
|
| 171 |
+
It is a modified copy of [PrivacyBench v2.5](https://huggingface.co/datasets/TonicAI/Privacy-Bench) by TonicAI, used
|
| 172 |
+
under CC BY 4.0. Modifications by Tracer: revised and added labels, spans marked unresolved, and an added file of
|
| 173 |
+
indirect identifiers. PrivacyBench+ is not affiliated with or endorsed by TonicAI.
|
| 174 |
+
|
| 175 |
+
## Citation
|
| 176 |
+
|
| 177 |
+
```bibtex
|
| 178 |
+
@misc{privacybenchplus2026,
|
| 179 |
+
title = {PrivacyBench+},
|
| 180 |
+
author = {Tracer},
|
| 181 |
+
year = {2026},
|
| 182 |
+
url = {https://huggingface.co/datasets/Tracer-ML/PrivacyBench-Plus}
|
| 183 |
+
}
|
| 184 |
+
```
|
| 185 |
+
|
| 186 |
+
Please also cite the source dataset, [TonicAI PrivacyBench](https://huggingface.co/datasets/TonicAI/Privacy-Bench).
|
SCHEMA-EXTENSIONS.md
ADDED
|
@@ -0,0 +1,50 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Schema extensions
|
| 2 |
+
|
| 3 |
+
PrivacyBench+ keeps the ten entity types and the `ground_truth_spans` format of PrivacyBench v2.5, so existing
|
| 4 |
+
scorers and saved predictions remain compatible. The extensions below are additive.
|
| 5 |
+
|
| 6 |
+
## Direct and indirect identifiers
|
| 7 |
+
|
| 8 |
+
Anonymization standards such as ISO/IEC 20889 and ISO/IEC 27559 distinguish two kinds of identifier:
|
| 9 |
+
|
| 10 |
+
- **Direct identifiers** point to a person or account on their own.
|
| 11 |
+
- **Indirect identifiers** name nobody, but narrow down who is involved when combined: an employer, a client, a
|
| 12 |
+
product only one company makes, a site, a city.
|
| 13 |
+
|
| 14 |
+
| Label | Identifier class |
|
| 15 |
+
|---|---|
|
| 16 |
+
| `NAME_GIVEN`, `NAME_FAMILY`, `EMAIL_ADDRESS`, `USERNAME`, `PHONE_NUMBER`, `ACCOUNT_NUMBER`, `EMPLOYEE_ID`, `LOCATION_ADDRESS`, `URL` | Direct |
|
| 17 |
+
| `ORGANIZATION` | Indirect |
|
| 18 |
+
|
| 19 |
+
## Indirect identifier file
|
| 20 |
+
|
| 21 |
+
`data/extensions/indirect_identifiers.jsonl` has one JSON object per span that is not in `ground_truth_spans` but
|
| 22 |
+
reveals an organization or a place.
|
| 23 |
+
|
| 24 |
+
| Field | Description |
|
| 25 |
+
|---|---|
|
| 26 |
+
| `file`, `line` | Data file and line number of the record |
|
| 27 |
+
| `source_id` | Stable record identifier |
|
| 28 |
+
| `start`, `end`, `text` | Character span in the record text |
|
| 29 |
+
| `kind` | `organization_indicator` (product, brand, drug, aircraft model, programme, ticker, abbreviation, or a named division or site of one company) or `location` (standalone city, site or venue) |
|
| 30 |
+
| `identifier_class` | Always `indirect` |
|
| 31 |
+
| `origin` | `excluded_label` (previously labelled as an organization), `declined_addition` or `declined_address` |
|
| 32 |
+
| `basis`, `reviewer`, `finding_id` | Reason for the entry and review reference |
|
| 33 |
+
|
| 34 |
+
Examples: "737" (Boeing), "SkyBonus" (Delta), "Downy" (Procter & Gamble), "See & Spray" (John Deere), "Mounjaro"
|
| 35 |
+
(Eli Lilly), "Retail Link" (Walmart), "WMT" (Walmart), "Amsterdam", "Denver hub".
|
| 36 |
+
|
| 37 |
+
Generic departments, internal tools, public bodies and chat slang are not indirect identifiers.
|
| 38 |
+
|
| 39 |
+
## Scores
|
| 40 |
+
|
| 41 |
+
- **PII detection:** the ten entity types in `ground_truth_spans`. Unresolved spans and indirect identifiers are
|
| 42 |
+
neutral: masking them is neither rewarded nor penalised.
|
| 43 |
+
- **Anonymization grade:** the labels and the indirect identifiers are both required. Reports character precision,
|
| 44 |
+
recall and F1 on the combined set, the share of indirect identifiers fully hidden, and the share of records with at
|
| 45 |
+
least one required span left visible.
|
| 46 |
+
|
| 47 |
+
## Coverage
|
| 48 |
+
|
| 49 |
+
The indirect identifier file covers organization-revealing names and standalone places. Job titles, dates and rare
|
| 50 |
+
events are not covered. Organization roles and sensitive content categories are not annotated in this version.
|
SCORECARD.md
ADDED
|
@@ -0,0 +1,59 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Scorecard
|
| 2 |
+
|
| 3 |
+
Results on the recommended evaluation set: the human-annotated records of `megan_donovan_eli_lilly` and
|
| 4 |
+
`nora_caterpillar`. Scores are computed with `score.py` on masked characters, with entity type ignored.
|
| 5 |
+
|
| 6 |
+
| Column | Definition |
|
| 7 |
+
|---|---|
|
| 8 |
+
| Precision, Recall, F1 | PII detection, on the 1,877 records without unresolved spans |
|
| 9 |
+
| Not fully hidden | Share of labelled spans with at least one character left visible |
|
| 10 |
+
| Anonymization-grade F1 | Labels and indirect identifiers both required, on all 1,962 records (112 indirect identifiers) |
|
| 11 |
+
| Indirect identifiers hidden | Share of indirect identifiers fully masked |
|
| 12 |
+
| Records with a leak | Share of the 1,488 records with something to hide in which a required span is partly visible |
|
| 13 |
+
|
| 14 |
+
The configuration column gives the label set or prompt used for each system. Open models were run with the entity
|
| 15 |
+
types listed; other configurations may give different results.
|
| 16 |
+
|
| 17 |
+
## Token classifiers and span models
|
| 18 |
+
|
| 19 |
+
| System | Configuration | Precision | Recall | F1 | Not fully hidden % | Anonymization-grade F1 | Indirect identifiers hidden % | Records with a leak % |
|
| 20 |
+
|---|---|---|---|---|---|---|---|---|
|
| 21 |
+
| **PIIX-560m** | default | 96.5 | 97.2 | 96.9 | 4.0 | 96.2 | 4.5 | 13.8 |
|
| 22 |
+
| **PIIX-560m** | `recall=4` | 92.1 | 98.5 | 95.2 | 2.5 | 94.7 | 18.8 | 9.2 |
|
| 23 |
+
| **PIIX-560m** | `recall=8` | 82.0 | 99.4 | 89.8 | 1.3 | 89.6 | 58.0 | 5.3 |
|
| 24 |
+
| Fastino GLiNER2.5-base | 10 benchmark types | 89.3 | 83.4 | 86.2 | 20.6 | 85.9 | 22.3 | 44.8 |
|
| 25 |
+
| Fastino GLiNER2-multi | 10 benchmark types | 85.2 | 84.4 | 84.8 | 18.6 | 83.6 | 48.2 | 42.3 |
|
| 26 |
+
| Fastino GLiNER2.5-multi | 10 short queries | 79.8 | 84.1 | 81.9 | 20.4 | 82.6 | 17.0 | 41.4 |
|
| 27 |
+
| Fastino GLiNER2-large | 10 benchmark types | 78.5 | 85.5 | 81.8 | 12.6 | 81.6 | 42.9 | 29.1 |
|
| 28 |
+
| Perplexity PII-Tracer | native labels | 86.1 | 72.9 | 79.0 | 26.7 | 78.2 | 8.9 | 56.5 |
|
| 29 |
+
| GLiNER multi-PII | 53 published types | 72.8 | 85.8 | 78.7 | 16.7 | 79.7 | 8.0 | 36.0 |
|
| 30 |
+
| Fastino GLiNER2.5-multi | 10 benchmark types | 85.7 | 69.3 | 76.6 | 41.2 | 77.7 | 9.8 | 61.2 |
|
| 31 |
+
| OpenAI Privacy Filter | native labels | 87.6 | 64.1 | 74.1 | 33.7 | 72.2 | 17.0 | 62.6 |
|
| 32 |
+
| Fastino GLiNER2 privacy-filter PII multi | 42 native labels | 70.6 | 75.6 | 73.0 | 23.1 | 73.0 | 42.9 | 45.0 |
|
| 33 |
+
| Fastino GLiNER2 Guardrails PII Multi | 42 native labels | 67.2 | 77.6 | 72.0 | 21.5 | 71.9 | 52.7 | 41.2 |
|
| 34 |
+
| NuMind NuNER Zero | 16 short queries | 61.4 | 79.7 | 69.4 | 25.1 | 71.0 | 31.2 | 22.2 |
|
| 35 |
+
| OpenMed privacy-filter-nemotron | 55 native labels | 66.8 | 72.0 | 69.3 | 30.4 | 68.9 | 1.8 | 59.5 |
|
| 36 |
+
| GLiNER large v2.1 | 16 short queries | 59.7 | 81.6 | 68.9 | 21.5 | 70.2 | 38.4 | 51.1 |
|
| 37 |
+
| Liquid LFM2.5 Encoder 350M PII | 40 native labels | 91.6 | 54.3 | 68.2 | 58.6 | 67.8 | 1.8 | 71.6 |
|
| 38 |
+
| PIIBench DeBERTa base | 82 native labels | 53.1 | 94.5 | 68.0 | 10.4 | 68.7 | 9.8 | 27.1 |
|
| 39 |
+
| Knowledgator GLiNER PII-large v1.0 | 40 published types | 77.3 | 56.7 | 65.4 | 48.3 | 63.1 | 28.6 | 53.4 |
|
| 40 |
+
| Fastino GLiNER2.5-multi | 28 ai4privacy types | 77.4 | 56.3 | 65.2 | 56.3 | 64.7 | 65.2 | 73.3 |
|
| 41 |
+
| Microsoft Presidio 2.2.364 | English, spaCy en_core_web_lg | 52.6 | 83.7 | 64.6 | 16.3 | 65.8 | 43.8 | 44.8 |
|
| 42 |
+
| Fastino GLiNER2.5-base | 28 ai4privacy types | 68.3 | 60.5 | 64.1 | 48.3 | 63.9 | 68.8 | 82.5 |
|
| 43 |
+
| Davlan XLM-R base NER HRL | 4 native labels | 74.8 | 56.1 | 64.1 | 36.3 | 66.2 | 1.8 | 77.8 |
|
| 44 |
+
| NVIDIA GLiNER-PII | 55 native labels | 52.1 | 79.3 | 62.9 | 21.1 | 62.4 | 27.7 | 37.7 |
|
| 45 |
+
| Knowledgator GLiNER bi-large v1.0 | 16 short queries | 46.5 | 95.6 | 62.6 | 10.7 | 64.6 | 93.8 | 33.8 |
|
| 46 |
+
| GLiNER multi v2.1 | 16 short queries | 60.5 | 64.1 | 62.2 | 31.0 | 62.6 | 9.8 | 58.9 |
|
| 47 |
+
| Piiranha | 17 native labels | 75.4 | 49.8 | 60.0 | 59.5 | 61.2 | 1.8 | 69.4 |
|
| 48 |
+
| Fastino GLiNER2-multi | 28 ai4privacy types | 53.6 | 67.3 | 59.7 | 37.8 | 59.5 | 27.7 | 55.3 |
|
| 49 |
+
| Fastino GLiNER2-large | 28 ai4privacy types | 51.1 | 69.7 | 59.0 | 29.1 | 58.8 | 50.9 | 48.6 |
|
| 50 |
+
| Gretel GLiNER bi-large v1.0 | 42 published types | 88.3 | 37.3 | 52.4 | 67.2 | 50.2 | 3.6 | 71.9 |
|
| 51 |
+
| SpanMarker mBERT MultiNERD | 15 native labels | 64.2 | 39.7 | 49.0 | 54.1 | 51.7 | 1.8 | 89.7 |
|
| 52 |
+
|
| 53 |
+
## Large language models
|
| 54 |
+
|
| 55 |
+
| System | Configuration | Precision | Recall | F1 | Not fully hidden % | Anonymization-grade F1 | Indirect identifiers hidden % | Records with a leak % |
|
| 56 |
+
|---|---|---|---|---|---|---|---|---|
|
| 57 |
+
| GPT-6 Astra | prompted | 99.4 | 96.9 | 98.1 | 6.4 | 97.8 | 0.9 | 22.2 |
|
| 58 |
+
| GPT-6.1 Sol | prompted | 99.3 | 96.8 | 98.1 | 6.6 | 97.7 | 0.9 | 22.4 |
|
| 59 |
+
| DeepSeek V4.1 Flash | span-extraction prompt | 79.3 | 93.4 | 85.7 | 9.0 | 85.6 | 95.5 | 22.2 |
|
data/extensions/indirect_identifiers.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/ground_truth/aaron_pfizer/spans.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/ground_truth/camille_nike/spans.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/ground_truth/daniel_lincoln_financial_group/spans.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/ground_truth/devon_marriott_international/spans.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/ground_truth/elena_stripe/spans.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/ground_truth/grace_deere_company/spans.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/ground_truth/hannah_salesforce/spans.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/ground_truth/isabella_chipotle/spans.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/ground_truth/malik_spotify/spans.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/ground_truth/marcus_boeing/spans.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/ground_truth/megan_donovan_eli_lilly/spans.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/ground_truth/morgan_morgan_stanley/spans.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/ground_truth/nora_caterpillar/spans.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/ground_truth/priya_charles_schwab/spans.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/ground_truth/ray_raymond_james/spans.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/ground_truth/renee_tyson_foods/spans.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/ground_truth/sofia_netflix/spans.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/ground_truth/tanya_kroger/spans.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/ground_truth/victor_delta_air_lines/spans.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/ground_truth/wells_wells_fargo/spans.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/ground_truth/yusuf_procter_gamble/spans.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/human_annotations/daniel_lincoln_financial_group.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/human_annotations/daniel_lincoln_financial_group_documents.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/human_annotations/isabella_chipotle.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/human_annotations/isabella_chipotle_documents.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/human_annotations/megan_donovan_eli_lilly.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/human_annotations/megan_donovan_eli_lilly_documents.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/human_annotations/morgan_morgan_stanley.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/human_annotations/morgan_morgan_stanley_documents.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/human_annotations/nora_caterpillar.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/human_annotations/nora_caterpillar_documents.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/human_annotations/renee_tyson_foods.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/human_annotations/renee_tyson_foods_documents.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
review/decisions.jsonl.gz
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d18354857796387a26ac016fe9d50e2e0685baf200c31c99e30d154d5616bf0e
|
| 3 |
+
size 1685260
|
score.py
ADDED
|
@@ -0,0 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Score a PII detector on PrivacyBench+. Standard library only.
|
| 2 |
+
|
| 3 |
+
python score.py DATA_DIR predictions.jsonl [--companies a,b] (default: the two suggested test companies)
|
| 4 |
+
|
| 5 |
+
predictions.jsonl: one JSON object per line, {"id": "<file name>:<line number>", "spans": [{"start": 0, "end": 5}, ...]}
|
| 6 |
+
where <file name> is the file under DATA_DIR/human_annotations and <line number> starts at 1. Entity type is ignored:
|
| 7 |
+
the benchmark measures whether sensitive characters are masked.
|
| 8 |
+
|
| 9 |
+
Two scores are printed.
|
| 10 |
+
PII detection labels in ground_truth_spans. Records with unresolved ("unknown") spans are left out.
|
| 11 |
+
Characters of indirect identifiers are neutral (masking them is neither rewarded nor punished).
|
| 12 |
+
Anonymization grade every record; labels and indirect identifiers are both required; unresolved spans are neutral.
|
| 13 |
+
"""
|
| 14 |
+
import json, sys
|
| 15 |
+
from collections import Counter, defaultdict
|
| 16 |
+
from pathlib import Path
|
| 17 |
+
|
| 18 |
+
TEST_COMPANIES = ['megan_donovan_eli_lilly', 'nora_caterpillar']
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
def bits(spans):
|
| 22 |
+
v = 0
|
| 23 |
+
for a, b in spans: v |= ((1 << (b - a)) - 1) << a
|
| 24 |
+
return v
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
def main():
|
| 28 |
+
data, pred_path = Path(sys.argv[1]), sys.argv[2]
|
| 29 |
+
companies = sys.argv[sys.argv.index('--companies') + 1].split(',') if '--companies' in sys.argv else TEST_COMPANIES
|
| 30 |
+
ind = defaultdict(list)
|
| 31 |
+
for l in (data / 'extensions/indirect_identifiers.jsonl').open():
|
| 32 |
+
t = json.loads(l)
|
| 33 |
+
if t['file'].startswith('human_annotations/'): ind[f"{Path(t['file']).name}:{t['line']}"].append((t['start'], t['end']))
|
| 34 |
+
recs = {}
|
| 35 |
+
for f in sorted((data / 'human_annotations').glob('*.jsonl')):
|
| 36 |
+
if f.stem.removesuffix('_documents') not in companies: continue
|
| 37 |
+
for n, line in enumerate(f.open(), 1):
|
| 38 |
+
r = json.loads(line)
|
| 39 |
+
recs[f'{f.name}:{n}'] = ([(s['start'], s['end']) for s in r['ground_truth_spans']],
|
| 40 |
+
[(u['annotation']['start'], u['annotation']['end']) for u in r.get('_tracer_review', {}).get('unknown_targets', [])])
|
| 41 |
+
pred = {}
|
| 42 |
+
for l in open(pred_path):
|
| 43 |
+
if l.strip(): p = json.loads(l); pred[p['id']] = [(s['start'], s['end']) for s in p['spans']]
|
| 44 |
+
c = Counter()
|
| 45 |
+
for rid, (gold, unknown) in recs.items():
|
| 46 |
+
if rid not in pred: c['missing'] += 1
|
| 47 |
+
q = bits(pred.get(rid, [])); g = bits(gold); targets = ind.get(rid, []); im = bits(targets)
|
| 48 |
+
for a, b in targets:
|
| 49 |
+
m = ((1 << (b - a)) - 1) << a; c['ind'] += 1; c['ind_hidden'] += (m & q) == m
|
| 50 |
+
req = g | im; neutral = bits(unknown) & ~req
|
| 51 |
+
c['a_tp'] += (req & q).bit_count(); c['a_pred'] += (q & ~neutral).bit_count(); c['a_gold'] += req.bit_count()
|
| 52 |
+
need = gold + targets
|
| 53 |
+
c['req_rows'] += bool(need); c['leak_rows'] += any((((1 << (b - a)) - 1) << a) & ~q for a, b in need)
|
| 54 |
+
if unknown: continue
|
| 55 |
+
c['rows'] += 1
|
| 56 |
+
c['tp'] += (g & q).bit_count(); c['pred'] += (q & ~(im & ~g)).bit_count(); c['gold'] += g.bit_count()
|
| 57 |
+
for a, b in gold:
|
| 58 |
+
m = ((1 << (b - a)) - 1) << a; c['spans'] += 1; c['hidden'] += (m & q) == m
|
| 59 |
+
f = lambda a, b: 100 * a / b if b else float('nan')
|
| 60 |
+
out = {'companies': companies, 'records': len(recs), 'records_without_prediction': c['missing'],
|
| 61 |
+
'pii_detection': {'records_scored': c['rows'], 'precision': f(c['tp'], c['pred']), 'recall': f(c['tp'], c['gold']),
|
| 62 |
+
'f1': f(2 * c['tp'], c['pred'] + c['gold']), 'spans_not_fully_hidden_pct': 100 - f(c['hidden'], c['spans'])},
|
| 63 |
+
'anonymization_grade': {'records_scored': len(recs), 'precision': f(c['a_tp'], c['a_pred']), 'recall': f(c['a_tp'], c['a_gold']),
|
| 64 |
+
'f1': f(2 * c['a_tp'], c['a_pred'] + c['a_gold']), 'indirect_identifiers': c['ind'],
|
| 65 |
+
'indirect_identifiers_hidden_pct': f(c['ind_hidden'], c['ind']),
|
| 66 |
+
'records_with_a_leak_pct': f(c['leak_rows'], c['req_rows']), 'records_with_something_to_hide': c['req_rows']}}
|
| 67 |
+
print(json.dumps(out, indent=1))
|
| 68 |
+
|
| 69 |
+
|
| 70 |
+
if __name__ == '__main__':
|
| 71 |
+
main()
|