newmean-pii-dataset (public test)

17,058 synthetic texts in 54 languages with character-level labels for personal information (PII), built to test PII detectors on the kinds of text people actually paste into forms, chats, spreadsheets and AI assistants. It is the public test of newmean-pii-dataset, the data behind the newmean-pii models (@newmean-ai/pii, browser-side masking).

• Synthetic. Every text was generated by Claude Sonnet 5.5 or Claude Opus 5.5; no row comes from a real person. Values were steered toward fictional ones, but not every value is checked, so we do not claim that all values are fictional. Write to hoonjae@newmean.ai to have any row removed. Version 1.1 replaced values that could match real ones (see Versions).
• Reviewed. Every row was re-tagged blind by a second model (details below).
• Private twin. The private test of the same design (17,001 rows, its own seeds) is kept unreleased. The leaderboard reports scores on both; on its first run every model scored within about one point of its public miss rate.
• Not for training. Please don't train or tune on this set; it is meant for evaluation.

Contents

rows	17,058 (305–320 per language)
rows without PII	6,862 (40%, hard negatives from the same domains and formats)
PII spans	17,147 (level 1: 15,775, level 2: 1,372)
characters	3.59 M
generators	Claude Sonnet 5.5 (9,686 rows; all SECRET values), Claude Opus 5.5 (7,372 rows)
design axes	54 languages × 16 domains × 20 text formats × 8 writing voices, plus length and twists (obfuscated values, mixed scripts, repeated mentions)

A row is one text a person types, pastes or uploads: a spreadsheet cell, a form or survey answer, a chat thread, an e-mail, a log line, an OCR excerpt — not a single sentence.

Fields

{"id": "pmt-00000", "text": "…", "lang": "ar",
 "spans": [{"start": 37, "end": 52, "label": "FINANCIAL", "level": 1, "subtype": "bic", "flags": ["obfuscated"]}],
 "domain": "banking, payments and insurance", "format": "OCR text from a scanned form, ID card or receipt (…)",
 "length": "short", "voice": "slang and abbreviations", "twist": "obfuscated or spoken-style PII (…)",
 "generator": "claude-sonnet-5-5", "review": "confirmed"}

• start/end: character offsets into text (Python string indices, i.e. Unicode code points), end exclusive. Spans never overlap.
• label: one of the labels below. subtype refines it (e.g. PERSON: full_name, surname, initials; PHONE: mobile, landline, toll_free); flags mark obfuscated (spelled out, spaced, "at"/"dot"), masked (partly hidden) and similar.
• level: 1 = directly identifying; 2 = indirect identifiers (a surname with a title, a nickname, initials, a public profile or page URL). Detectors that only target level 1 can be scored on level 1 alone.
• review: confirmed (the blind reviewer reproduced the labels), changed (both reviewers agreed on a change), adjudicated.

Labels

label	spans	covers
PERSON	5,878	names of people, incl. given names, surnames with titles, nicknames, initials (the last three are level 2)
ID_NUMBER	2,691	national IDs, passports, driver's licences, tax, student, employee, patient and respondent numbers, vehicle and device identifiers
PHONE	1,980	phone, mobile, fax and full toll-free numbers
FINANCIAL	1,231	card, account and IBAN numbers, expiry/CVV, payment handles, crypto wallets, SWIFT/BIC
EMAIL	1,221	e-mail addresses
ADDRESS	1,129	addresses with street, building, unit or postal code; precise meeting points
ORG	902	organisations a specific person belongs to or is treated at (employer, school, clinic)
SECRET	620	passwords, one-time codes, API keys and tokens
USERNAME	612	handles and account names
DATE_OF_BIRTH	370	dates of birth (full or partial)
URL	330	personal URLs (level 1) and public profile or page URLs (level 2)
IP_ADDRESS	183	IP addresses

Not labelled: a city or country alone, roles and titles alone, brands and shops a person only uses, order, ticket and tracking numbers that do not point to a person, amounts, and ordinary dates.

How it was made

Texts were generated per item from a written specification (language and country, domain, format, length, voice, how many PII values, twists) with seed values from per-country pools, returned with inline tags and parsed into offsets by a strict parser. Rows failing automatic checks (spec coverage, format validity, language) were held out. The set was then deduplicated against the training, validation and private test data (exact texts and MinHash near-duplicate templates); rows are only removed from the lower-priority split, so no private test row can appear here. The generator code and prompts are not released; the technical report describes the method.

Review. Claude Opus 5.5 re-tagged all 17,058 rows from the raw text alone, without seeing the generated labels. It reproduced them on 16,730 rows (98.1%), where two independent annotations agree. The 328 rows where it differed were re-tagged blind by GPT-6.1 Sol as well (plus 4,400 matching rows in an early full pass): both reviewers agreed on a label change in 220 rows, the generated labels held by majority in 181, and the 25 open rows were adjudicated (GLM through ZCode chose among the versions; each choice was checked against the guideline, 2 overridden). In total 318 rows (1.9%) changed. Errors that the generator and the reviewer share remain possible.

Evaluation

The newmean metrics, from the masking point of view:

• leak: share of PII characters not covered by any predicted span (top priority; lower is better);
• overmask: share of non-PII characters that get masked;
• span F1 (overlap ≥ 50%, same label) and exact F1 (same boundaries and label).

Detectors with a different label set are scored after mapping their labels to these; leak and overmask do not depend on labels.

Results on version 1.1 (against 1.0, the newmean models moved by at most 0.05 points and the other detectors by 0.2–1.6). Model only, no regex layer, single runs on a Colab L4 VM; public detectors with their recommended settings, scored against labels collapsed to their own label set.

model	size	miss % (leak)	over-masking %	exact F1
newmean-pii-large	15.5 MB	1.6	0.6	0.78
newmean-pii-medium	11.3 MB	2.1	0.6	0.77
newmean-pii-small	7.1 MB	2.7	1.2	0.68
newmean-pii-nano	2.6 MB	3.3	2.1	0.59
knowledgator/gliner-pii-edge-v1.0	176 MB	19.6	31.9	0.09
openai/privacy-filter	2.8 GB	34.6	4.4	0.25
Horizon-Labs/pii-redactor-small	563 MB	41.0	1.1	0.37
atonlee/koelectra-ko-pii-ner	14M params	43.1	10.2	0.12
fastino/gliner2-privacy-filter-PII-multi	1.2 GB	44.9	1.6	0.52
nationaldesignstudio/rampart	15 MB	56.5	3.0	0.24
Mozilla/smart-pii-redactor-v1-onnx	15 MB	56.9	2.0	0.28
nationaldesignstudio/rampart-edge	1.5 MB	74.7	1.5	0.17
regex baseline	—	75.3	0.6	0.18

The newmean models were trained on data of the same design and guideline (generated by Claude Haiku and Sonnet), so they have an in-distribution advantage; several public detectors intentionally leave some labels (e.g. organisations) unmasked. Per-label, per-language and per-level numbers and the conditions of the comparison are in the technical report (Section 6, Appendix O).

Versions

• 1.1 (2026-10-31): a scan of every value found ones that could match real people or accounts, and they were replaced in 569 rows (3.3%): 479 free-mail e-mail addresses moved to reserved domains (RFC 2606/6761), checksum-valid national ID numbers (76 rows), payment cards (17) and an IBAN (1) had one digit changed so the check fails, and one public figure's name was replaced by a fictional name. Labels and span counts are unchanged; offsets follow the new values. The row list is in CHANGES.json (ids and kinds only).
• 1.0 (2026-10-31): first release. Its files are no longer kept in this repository.

License and use

CC BY-NC 4.0: free for research and evaluation with attribution. For commercial use contact hoonjae@newmean.ai. The training and validation data are not on the Hub; request them through the data request form (requests open now; the data is shared from December 2026).

Made by newmean.ai.