LFM2.5-Encoder-350-PII-Detector
A full fine-tune of LFM2.5-Encoder-350M with a token-classification head, covering 40 PII types across 16 languages.
Entity types (40 PII types across 11 domains)
| Domain | Types |
|---|---|
| Identity | identity.person_name, identity.ssn, identity.national_id, identity.passport, identity.drivers_license, identity.date_of_birth, identity.tax_id |
| Contact | contact.email, contact.phone, contact.address, contact.postal_code, contact.ip_address |
| Financial | financial.credit_card, financial.iban, financial.bank_account, financial.swift_bic, financial.crypto_wallet, financial.amount |
| Credentials | credential.api_key, credential.password, credential.private_key, credential.jwt, credential.connection_string, developer.login_credentials |
| Online | online.username, online.url |
| Device | device.mac_address, device.imei, developer.device_id |
| Location | location.gps_coordinates |
| Healthcare | healthcare.medical_record, healthcare.condition, healthcare.medication, healthcare.health_plan_id |
| Organization | org.company_name |
| Special-category | special.religion, special.political, special.orientation, special.health_status |
| Legal | legal.case_number |
Benchmarks (18-locale-filtered, partial-F1, hybrid decode)
| Benchmark | this model | detection tier | SauerkrautLM GLiNER | openai/privacy-filter | Piiranha-v1 | OpenMed privacy-filter | regex + validators |
|---|---|---|---|---|---|---|---|
| SPY | 0.428 | 0.509 | 0.280 | 0.264 | 0.232 | 0.226 | 0.358 |
| Gretel | 0.880 | 0.885 | 0.663 | 0.458 | 0.553 | 0.770 | 0.337 |
| TAB | 0.867 | 0.888 | 0.685 | 0.543 | 0.262 | 0.672 | 0.000 |
| ai4privacy | 0.715 | 0.774 | 0.488 | 0.394 | 0.946 | 0.432 | 0.195 |
| Nemotron | 0.855 | 0.863 | 0.639 | 0.572 | 0.658 | 0.918 | 0.335 |
| MAPA | 0.236 | 0.267 | 0.416 | 0.288 | 0.228 | 0.164 | 0.000 |
- Best on every benchmark except MAPA, whose idiosyncratic date-as-
date_of_birthlabeling convention penalises correctly-typed predictions. Only two external scores land higher anywhere — Piiranha-v1's 0.946 on ai4privacy and OpenMed's 0.918 on Nemotron — and both are in-distribution: each model trains on that exact corpus, as did our own encoder's pretraining on those same two. - Detection tier is the same model and the same predictions, scored with the type label ignored — did it find the PII span at all, which is the metric that matters for redaction. The gap to exact-type is the type-confusion rate, e.g. SPY 0.428 → 0.509.
Usage
⚠️ Loads custom code via
trust_remote_code=True(the model wraps atrust_remote_codeencoder).
Install the required packages:
pip install torch transformers huggingface_hub
Run PII detection:
import importlib.util
import sys
from huggingface_hub import hf_hub_download
from transformers import AutoModelForTokenClassification, AutoTokenizer
model_id = "LiquidAI/LFM2.5-Encoder-350-PII-Detector"
helper_path = hf_hub_download(model_id, "pii_hybrid_decode.py")
hf_hub_download(model_id, "context_cued.py")
sys.path.insert(0, helper_path.rsplit("/", 1)[0])
spec = importlib.util.spec_from_file_location("pii_hybrid_decode", helper_path)
hd = importlib.util.module_from_spec(spec)
spec.loader.exec_module(hd)
tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForTokenClassification.from_pretrained(model_id, trust_remote_code=True).eval()
spans = hd.predict("Email Dr. Laura Schmidt at laura@charite.de.", tok, model)
print(spans)