abubasith86/pii-dataset-de-v1

PII Datasets v1 (German)

German training and evaluation data for the PII detection model. Same 18 labels, format and labelling guide as the English dataset abubasith86/pii-dataset-v1 and the Arabic dataset abubasith86/pii-dataset-ar-v1.

Scope

• Text: German business and everyday text from Germany, Austria and Switzerland: support messages, emails, forms.
• Task: find each PII item, give its type and character position.
• Data: public datasets only (no data generated by us).

Labels

The 18 labels of the English dataset; see labels.json. What to mark: every value of the 18 labels, company ones too (labelling_rules.txt).

Format

JSON Lines, one document per line; locale is de, de-DE, de-AT or de-CH:

{"id": "kiji_de:train-1234", "source": "kiji_de", "domain": "generic", "locale": "de-DE",
 "text": "Hallo, ich bin Jana Weber, meine IBAN ist DE44 5001 0517 5407 3249 31.",
 "masked_text": "Hallo, ich bin [PERSON_NAME], meine IBAN ist [BANK_ACCOUNT].",
 "spans": [{"start": 15, "end": 25, "label": "PERSON_NAME"}, {"start": 42, "end": 69, "label": "BANK_ACCOUNT"}]}

start and end are character positions (end not included). An empty spans list means the document has no PII.

Sources

Source	Dataset	What is used
ai4privacy_de	ai4privacy/pii-masking-openpii-1.5m	German rows (DE, AT, CH); up to 40,000 train docs; its validation file is split 50/50 into dev and test
nym_de	Wismut/nym-pii-multilingual-data (MIT)	German rows (Latin-1 text with German function words); its train / validation / test splits
kiji_de	DataikuNLP/kiji-pii-training-data (Apache-2.0)	German rows (Germany, Austria, Switzerland); 5% of its train split is dev
ai4privacy400k_de	ai4privacy/pii-masking-400k	German files; up to 20,000 train docs; validation split 50/50 into dev and test; no text shared with openpii-1.5m
ai4privacy300k_de	ai4privacy/pii-masking-300k	German files; up to 15,000 train docs; documents whose span text differs from its value, or with template slots left in the text, are skipped
nobody_de	naeyn/nobody-pii-synth-de (Apache-2.0)	German rows: rule-generated tickets, emails, letters and chats; token positions converted to character positions

Not used for training: gretelai/synthetic_pii_finance_multilingual German files (common nouns tagged as names, IBANs and emails left unlabelled) and E3-JSI/synthetic-multi-pii-ner-v1 German rows: both are kept as outside test sets.

Checks

• Every source was audited before use: span text matches its character positions; unlabelled emails, German IBANs and phone numbers were searched for.
• ai4privacy puts random codes behind ID labels: a document whose national ID, passport, driving licence or tax ID has no matching word before it (Personalausweis, Reisepass, Führerschein, Fahrerlaubnis, Steuer-ID, Identifikationsnummer, ...) is dropped.
• kiji labels only the first mention of a value: later mentions get the same label.
• A number alone tagged as an address ("9 g", "Raum 301"), a label word tagged as a value ("IBAN", "Führerschein") and IDs of only 1-3 letters ("Ausweis: JAK") are removed.
• The same cleaning as the English dataset: exact and near duplicates removed (same text with other PII values), train/dev texts that repeat a test text removed, every date, email, URL and phone labelled (company ones too), titles (Herr, Frau, Dr.) trimmed off names, fake-looking ID digits (1234, 5555) replaced.
• A document whose locale is German must not be mostly English.

Statistics

Source	train	dev	test
ai4privacy_de	29,360	2,236	3,674
nym_de	9,811	522	787
kiji_de	7,106	384	852
ai4privacy400k_de	17,052	1,282	2,120
ai4privacy300k_de	11,404	747	1,480
nobody_de	4,659	270	279
Total	79,392	5,441	9,192

Domain	train	dev	test
generic	79,392	5,441	9,192