ApplePie

Open PII detection model built for Europe. By The General Data Corporation.

ApplePie finds personal data in text across European languages: names, addresses, national ID numbers, IBANs, payment cards, phone numbers, emails, dates of birth, health and customer identifiers, licence plates and more. It recognises the real formats of European identifiers, including French NIR and SIRET, German Steuer-ID, Spanish DNI and NIE, Italian codice fiscale, Dutch BSN, Polish PESEL, Portuguese NIF, Belgian national numbers and IBANs.

On real European legal text in 21 languages, ApplePie reaches 0.846 F1, against 0.546 for the strongest open detector we measured.

Results

F1 on held-out test sets. Higher is better.

Benchmark	What it is	ApplePie	NVIDIA GLiNER-PII	Microsoft Presidio	GLiNER multi-PII	Gretel GLiNER
MAPA	Real EU legal text, 21 languages	0.846	0.546	0.166	0.490	0.296
GD Europe	Business records, 9 EU countries	0.997	0.876	0.605	0.853	0.525
Nemotron-PII	NVIDIA, English	0.975	0.830	0.713	0.749	0.448
Gretel PII	Gretel, English	0.997	0.905	0.721	0.840	0.826
TAB	Court rulings, English	0.719	0.709	0.739	0.594	0.418

Precision and recall for ApplePie:

Benchmark	Recall	Precision	F1
MAPA	0.813	0.881	0.846
GD Europe	0.999	0.995	0.997
Nemotron-PII	0.966	0.984	0.975
Gretel PII	0.998	0.995	0.997
TAB	0.604	0.889	0.719

How these were measured

• Each system is scored on the first 1,000 documents of each test split (all 127 for TAB). A personal data span counts as found when at least half of its characters are covered. This is our own scoring rule, not each benchmark's official metric.
• Open models run at their default settings: GLiNER models at a 0.5 threshold, Presidio at 0.4. ApplePie runs at its default 0.5.
• ApplePie was trained on the training splits of MAPA, Nemotron-PII, Gretel and GD Europe, and scored only on their held-out test splits. That gives it a home advantage on those four, as NVIDIA's model has on Nemotron-PII. ApplePie never saw TAB.
• GD Europe is our own benchmark, generated by the same process as ApplePie's European training data, from separate templates. MAPA is the independent European test: real text, annotated by the EU-funded MAPA project.
• No setting was tuned on any test set.

Usage

from transformers import pipeline

pii = pipeline("token-classification", model="generaldata/ApplePie", aggregation_strategy="simple")

for found in pii("Bonjour, je suis Marie Dubois, née le 03/08/1980. Mon IBAN est FR76 3000 6000 0112 3456 7890 189."):
    print(found["entity_group"], found["start"], found["end"], round(found["score"], 3))

Each result carries the label (entity_group), the character offsets (start, end), the matched text (word) and a confidence (score).

Texts longer than 512 tokens should be split into overlapping windows; the pipeline's stride argument does this.

Labels

ApplePie tags 20 kinds of personal data, in BIO format.

Label	Covers
person_name	Full names, first names, surnames
organization	Companies and institutions
street_address	Street addresses and postcodes
location	Cities, regions, countries
coordinate	GPS coordinates
date	Dates and times
date_of_birth	Dates of birth
email	Email addresses
phone_number	Phone and fax numbers
url	Personal links and profiles
user_name	Logins and usernames
card_number	Payment card numbers and CVVs
bank_account	IBANs, account and routing numbers, SWIFT/BIC
national_id	National IDs, social security and tax numbers, passports, licences
secret	Passwords, PINs, API keys, cookies
device_identifier	IP and MAC addresses, device IDs
customer_id	Customer, contract and employee numbers
health_id	Medical record, health insurance and biometric identifiers
vehicle_identifier	Licence plates and VINs
trait	Age, gender, occupation, nationality and other personal attributes

Training

• Base model: XLM-RoBERTa large, multilingual.
• Data:
  • Nemotron-PII training split (NVIDIA)
  • Gretel PII masking training split (Gretel)
  • MAPA training split: EU legal text in 21 languages
  • GD Europe, our own corpus of 17,950 business documents from France, Germany, Spain, Italy, the Netherlands, Portugal, Poland, Belgium, Ireland and the UK. The documents were written by an open model, Mistral Nemo, with placeholders for personal data. Each placeholder was then filled with values in that country's real format, so every label is exact and every identifier carries valid check digits.
• Setup: one epoch, learning rate 2e-5, batch size 16, sequence length 512, bf16, on one NVIDIA A100.

Limitations

• Detection is not anonymisation. Under the GDPR, pseudonymised data is still personal data; use ApplePie as one control among several, with human review for sensitive releases.
• Much of the European training data is synthetic. Real documents in unfamiliar layouts or domains may score lower than the benchmarks above.
• Context-dependent personal data, such as a rare job title that identifies one person, is harder to catch than structured identifiers. On TAB, a benchmark built around this, ApplePie is close to but below Presidio.
• Languages outside those listed above are untested.

License

ApplePie is © 2026 The General Data Corporation and released under Creative Commons Attribution-NonCommercial 4.0 (CC BY-NC 4.0).

• Non-commercial use (research, evaluation, personal projects) is free, with attribution: "ApplePie by The General Data Corporation".
• Commercial use requires a commercial licence from The General Data Corporation. Contact us at generaldata.io.

See LICENSE for the full terms and third-party notices.

Citation

@misc{applepie2026,
  title        = {ApplePie: PII detection built for Europe},
  author       = {Allani, Louai and {The General Data Corporation}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/generaldata/ApplePie}}
}

Author

Louai Allani, The General Data Corporation.