naeyn/nobody-pii-synth-de

Nobody PII Synthetic

A procedurally generated, multilingual corpus for training and regression testing PII detection and redaction systems, with German as the primary language and English/Dutch coverage.

Companion model: naeyn/nobody-pii-de

Release identity

This Hub revision contains the canonical Parquet snapshot. The Parquet files were verified at Hub revision 23eba3fadef50ef4c01738bea0b0ee1c392426e6; card and attribution metadata may receive later commits without changing these data files. Pin the data files by revision and verify their hashes when exact bytes matter.

file	SHA-256
data/train.parquet	19ac658494c9e7d1e835afe257d4424e6aea775f32322c1b68d9852dc59fe0d8
data/validation.parquet	447da873078dae51b6685c23c6d8a2a1d8375ffc923267108efed216eaaa6b1a
data/test.parquet	578ce3d52b621dec093be0e685d85183ffa7f238756f3d59b4aa18a865ab0919

Data statement

This release was generated from templates and Faker locale providers. No source records containing real people's personal data were used to construct these splits. Names, addresses, dates, contact details, identifiers, and other values are generated or randomly constructed for the corpus. Any resemblance to a real person, organization, or address is coincidental and unintended.

The dataset does not contain records from ai4privacy/pii-masking-400k or ai4privacy/pii-masking-openpii-1m. Those are separate sources with separate licensing and provenance. OpenPII is used only as a documented training source for the companion model, not as data in this repository.

Generation provenance

• Generator snapshot: gen_v4.1 document-composer revision. It includes business-document templates, multilingual name forms, structured-identifier traps, ordinary-date hard negatives, pronoun/role-noun negatives, and Swiss address forms.
• Requested corpus size: 10,000 generated documents plus a 5% supplement of 500 pure-negative controls, for 10,500 rows total.
• Split construction: one shuffled corpus split into 9,450 train, 525 validation, and 525 test documents (90% / 5% / 5%). The source JSON names the middle split eval; the Hub export names it validation.
• Generation seed: not retained in the canonical JSON snapshot. The training mix used seed 42 for concatenation/shuffling, but that is a separate model-training seed and must not be presented as this corpus's generation seed.
• Generation date: not embedded in the canonical data artifact. The gen_v4.1 label identifies the generator revision, not a reproducible timestamp. The published Parquet files and hashes above are the authoritative snapshot.

A fresh run of the public generator family may reproduce the distribution and template family but is not claimed to reproduce these exact rows. This explicitly avoids implying byte-for-byte regeneration from metadata that was not recorded.

Splits and languages

split	documents	German	English	Dutch
train	9,450	5,349	2,265	1,836
validation	525	300	130	95
test	525	296	127	102
total	10,500	5,945	2,522	2,033

The held-out test split contains 3,426 labeled entity spans and is intended as regression evidence for models trained on this corpus. It is not an independent real-world benchmark.

Identity, template, and split policy

A generated document intentionally reuses one document-local identity in its header, body, signature, and related fields when the template calls for it. Documents are otherwise generated independently. No global identity registry, identity-level grouping, or cross-split uniqueness guarantee was implemented; Faker can therefore produce coincidental value collisions across documents. Template families and negative controls are allowed to recur in every split.

This is a transparent synthetic-regression design, not a privacy-safe claim that the held-out split is independent in the statistical sense.

Deduplication and split-contamination audit

An exact-text audit of all 10,500 canonical JSON rows found:

• 112 duplicate-text groups in the complete corpus;
• 36 groups spanning more than one split;
• 0 duplicate groups containing any labeled entity span.

The cross-split duplicate groups are therefore pure-negative boilerplate, not repeated labeled PII. This check is exact-text only; because templates recur, near-duplicate and semantic overlap are expected and are not represented by a claim of independence. The companion model's external German benchmark audit used a separate exact-hash and 5-gram-near-duplicate procedure, reported in that model card.

Features

Each Parquet row has these fields:

field	type	description
text	string	original generated document
tokenized_text	list[string]	whitespace/punctuation token sequence
ner	list[struct]	labeled token spans
lang	string	de, en, or nl
doctype	string	generated document template family

Each ner item is a struct with:

{"start": 5, "end": 5, "label": "person"}

start and end are inclusive token indices into tokenized_text, not character offsets. The public Parquet release uses typed structs so it can be loaded by the Hub Dataset Viewer and datasets without schema ambiguity.

Label taxonomy

• person — person-name mention
• address — postal or street address
• date of birth — birth date mention
• email — email address
• phone number — phone number
• iban — IBAN-shaped identifier
• organization — company or employer name

Dates used as delivery, invoice, contract, or other non-birth events are included as hard negatives when they are not labeled as date of birth. Structured identifiers are synthetic and may be format-valid; they do not represent accounts or identifiers belonging to real people.

Generation characteristics

The corpus contains business-style emails, chats, support tickets, letters, contracts, records, plain prose, and pure-negative controls. It deliberately includes:

• multilingual names and surname particles;
• punctuation, spacing, Unicode, and OCR/export variants;
• dates that require context to distinguish birth dates from ordinary dates;
• order, ticket, and customer-number traps;
• checksum-shaped synthetic identifiers; and
• exact labels produced at generation time rather than by a separate annotator.

Training split label counts

label	instances
person	24,974
email	10,472
address	7,865
organization	5,283
phone number	6,473
date of birth	3,409
iban	2,415

Intended use

Use this dataset to train or regression-test PII detection and redaction systems, especially for multilingual business-document workflows. Always measure on representative real or appropriately licensed evaluation data before making production, privacy, or compliance claims.

Limitations

• Template-based synthetic text cannot model the full diversity, noise, OCR, formatting, or ambiguity of real documents.
• Faker locale distributions are not a representation of all names, cultures, scripts, or geographies.
• Models can memorize templates or generator artifacts rather than learn transferable PII recognition.
• Checksum-shaped values are synthetic and must not be treated as live account or identity data.
• Exact byte-level regeneration is not promised because the original generation seed and timestamp were not preserved in the published snapshot.

Loading

from datasets import load_dataset

dataset = load_dataset(
    "naeyn/nobody-pii-synth-de",
    revision="23eba3fadef50ef4c01738bea0b0ee1c392426e6",
)
print(dataset)

License and attribution

This dataset is released under Apache-2.0. The generated rows do not include third-party source records. Faker locale providers were used as a generation tool; Faker is MIT-licensed.

The dataset is separate from the companion model's additional training source, ai4privacy/pii-masking-openpii-1m, which is used under CC-BY-4.0 with attribution to Ai4Privacy / Ai Suisse SA. No OpenPII rows are included here.

Citation

@dataset{nobody_pii_synth_de,
  title = {Nobody PII Synthetic},
  author = {naeyn},
  year = {2026},
  url = {https://huggingface.co/datasets/naeyn/nobody-pii-synth-de},
  license = {Apache-2.0}
}