pii-extractor-68M

Finds personal information (PII) in English text and returns the exact position of each value with one of 12 labels. 68M parameters, about 23 ms per 1,000 tokens on a T4 GPU.

Fine-tuned from jhu-clsp/ettin-encoder-68m. Current version: v0.1 (use --revision v0.1 to pin it). Also available in 32M and 17M sizes.

Labels

Label	Covers
PERSON	Names, including first names only, initials and nicknames
EMAIL	Email addresses
PHONE	Phone and fax numbers
ADDRESS	Street addresses (a city or country on its own is not labelled)
DATE	Dates of birth only
URL	Web addresses, with or without https://
ACCOUNT_NUMBER	Bank, card, IBAN and customer account numbers (not ticket or order numbers)
GOV_ID	Social security, passport, national ID, tax ID and driving licence numbers
USERNAME	Usernames and handles
SECRET	Passwords, PINs, access codes, API keys and tokens
IP_ADDRESS	IPv4 and IPv6 addresses
OTHER_ID	Employee, patient, member and device IDs

How to use

1. Download and install

pip install -U huggingface_hub
hf download CobuildX-AI/pii-extractor-68M --revision v0.1 --local-dir pii-extractor-68M
cd pii-extractor-68M
pip install -r requirements.txt

2. Use it in your code

PIIExtractor (in pii.py, part of this repo) loads the model once:

from pii import PIIExtractor

pii = PIIExtractor()                 # uses a GPU if one is available
pii.extract("Hi team, I'm Priya Raman (priya.raman@northwind.com).")
# [{'label': 'PERSON', 'text': 'Priya Raman', 'start': 13, 'end': 24},
#  {'label': 'EMAIL', 'text': 'priya.raman@northwind.com', 'start': 26, 'end': 51}]

pii.redact("Hi team, I'm Priya Raman (priya.raman@northwind.com).")
# "Hi team, I'm [PERSON] ([EMAIL])."

• The model finds the PII; redact() replaces each found value with its label (or your own template).
• Context length: 7,999 tokens, read in a single pass. Only longer texts are split, automatically.

3. Optional: check your setup

python example.py runs three sample texts and prints what it finds.

Results

Exact-match span F1 on two public benchmarks never used for training: a value counts only if its start, end and label are all correct.

Model	Size	Privacy-Bench	PII-TRACE	T4 GPU, 1,000 tokens
pii-extractor-68M	68M	88.3%	82.4%	23 ms
pii-extractor-32M	32M	87.5%	84.6%	16 ms
pii-extractor-17M	17M	86.0%	81.2%	15 ms
PII-Tracer	596M	86.5%	92.6%	271 ms
GLiNER2-PII	205M	83.7%	62.2%	199 ms

Privacy-Bench: 656 human-annotated business document pages. PII-TRACE: 500 conversations; it is PII-Tracer's own benchmark, and most of our gap there comes from dates and reference codes, which we deliberately don't label.

How we evaluated

We ran every model ourselves, on the same texts, with the same scoring. Privacy-Bench contains 656 pages of Word and PDF business documents with 8,490 PII values annotated by people. PII-TRACE contains 500 multi-turn conversations with 2,442 annotated values (public subset). Neither was used for training, and any training text that also appears in either benchmark was removed.

A prediction counts as correct only when its start, end and type all match the annotation. Each benchmark's types were mapped to ours as below, and every model was scored on the types that both the benchmark and the model support.

PII Extractor	Privacy-Bench	PII-TRACE
PERSON	NAME_GIVEN, NAME_FAMILY	private_person
EMAIL	EMAIL_ADDRESS	private_email
PHONE	PHONE_NUMBER	private_phone
ADDRESS	LOCATION_ADDRESS	private_address
URL	URL	private_url
ACCOUNT_NUMBER	ACCOUNT_NUMBER	account_number
USERNAME	USERNAME	Not annotated
DATE	Not annotated	private_date
SECRET	Not annotated	secret

GOV_ID, IP_ADDRESS and OTHER_ID are not annotated in either benchmark, so this evaluation does not cover them. Latency is the median of 50 texts of 1,000 tokens, one text at a time on a Tesla T4 GPU.

Limitations

• English only.
• Sensitive attributes such as health, religion or ethnicity are not detected in v0.1.
• Usernames, account numbers and links are found less reliably than names, emails and phones.
• A detection aid, not a guarantee: keep human review where redaction must be complete.

Licence

Apache-2.0. Trained on NVIDIA Nemotron-PII (CC-BY-4.0), Gretel PII masking EN v1 (Apache-2.0), Wismut nym-pii (MIT) and data we generated. Base model: Ettin (MIT, arXiv:2507.11412).