privacy-filter-nemotron-v2

OpenMed/privacy-filter-nemotron-v2 is the second-generation Nemotron-schema checkpoint in the OpenMed privacy-filter family. It keeps the same fine-grained 55-category PII vocabulary as OpenMed/privacy-filter-nemotron, while using a broader training mix and a more recall-oriented adaptation recipe. In practice, this v2 checkpoint should perform better as a general PII masking and redaction model while preserving the useful typed labels from the original Nemotron model.

The model is based on openai/privacy-filter, a 1.4B-parameter MoE token classifier with roughly 50M active parameters per token. It predicts 221 BIOES token classes:

• O
• 55 PII categories encoded as B-*, I-*, E-*, and S-*

Use this checkpoint when you want the Nemotron fine-grained label schema, but prefer the improved v2 masking behavior.

Relationship To The Original Nemotron Model

This model is a direct successor to OpenMed/privacy-filter-nemotron.

• Same base architecture: openai/privacy-filter
• Same core label schema: 55 fine-grained Nemotron-style PII categories
• Same output format: BIOES token classification
• Broader adaptation data: Nemotron-style fine labels plus additional PII masking examples from other synthetic PII sources
• Better practical masking behavior for general redaction use cases

The original OpenMed/privacy-filter-nemotron remains useful when you want the cleanest single-dataset Nemotron training lineage. This v2 model is the better default when you want stronger general-purpose PII masking while keeping the same fine-grained schema.

Quick Start

With OpenMed

pip install -U "openmed[hf]"

from openmed import extract_pii, deidentify

model_name = "OpenMed/privacy-filter-nemotron-v2"
text = (
    "Patient Sarah Johnson (DOB 03/15/1985), MRN 4872910, "
    "phone 415-555-0123, email sarah.johnson@example.com."
)

result = extract_pii(text, model_name=model_name)
for ent in result.entities:
    print(ent.label, ent.text)

masked = deidentify(text, method="mask", model_name=model_name)
print(masked.deidentified_text)

With opf

pip install 'opf @ git+https://github.com/openai/privacy-filter.git'

opf redact \
  --checkpoint OpenMed/privacy-filter-nemotron-v2 \
  --text "Patient Sarah Johnson (DOB 03/15/1985), MRN 4872910, phone 415-555-0123."

With Transformers

from transformers import AutoModelForTokenClassification, AutoTokenizer, pipeline

repo_id = "OpenMed/privacy-filter-nemotron-v2"

tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModelForTokenClassification.from_pretrained(
    repo_id,
    trust_remote_code=True,
)

ner = pipeline(
    "token-classification",
    model=model,
    tokenizer=tokenizer,
    aggregation_strategy="simple",
)

text = "Patient Sarah Johnson, MRN 4872910, can be reached at sarah@example.com."
print(ner(text))

For best production behavior, use BIOES-aware decoding and merge overlapping or consecutive spans before masking.

Label Space

The checkpoint uses 55 fine-grained PII categories:

• Identity and demographic attributes: first_name, last_name, age, gender, race_ethnicity, sexuality, religious_belief, political_view, marital_status, nationality, education_level, occupation, employment_status, language, blood_type, biometric_identifier
• Contact and web identifiers: email, phone_number, fax_number, url
• Address: street_address, city, county, state, country, postcode, coordinate
• Dates and times: date, date_of_birth, date_time, time
• Government and regulated IDs: ssn, national_id, tax_id
• Financial and secret values: account_number, bank_routing_number, swift_bic, credit_debit_card, cvv, pin, password
• Healthcare identifiers: medical_record_number, health_plan_beneficiary_number
• Enterprise and customer identifiers: customer_id, employee_id, unique_id, certificate_license_number
• Vehicle identifiers: license_plate, vehicle_identifier
• Digital identifiers: ipv4, ipv6, mac_address, device_identifier, api_key, http_cookie

The full label-space JSON is included as label_space_fine_v1.json.

Training Summary

This checkpoint was initialized from the first-generation OpenMed Nemotron privacy-filter branch and further adapted with source-balanced typed PII examples.

• Base model: openai/privacy-filter
• First-generation predecessor: OpenMed/privacy-filter-nemotron
• Output schema: 55 fine-grained PII labels, 221 BIOES classes
• Training precision: bf16
• Training method: full fine-tuning with OpenAI's opf train

The training mix includes synthetic PII examples derived from:

• nvidia/Nemotron-PII
• gretelai/gretel-pii-masking-en-v1
• ai4privacy/pii-masking-openpii-1m

Limitations And Intended Use

This is an experimental private checkpoint intended for PII detection, masking, and de-identification workflows. It should be validated on your target domain before use in high-stakes systems.

For clinical PHI, radiology/DICOM workflows, legal data, or other regulated settings, use this model as one component inside a broader de-identification pipeline with deterministic rules, audit logging, and human review where appropriate.

Credits

This model builds on:

• OpenAI's openai/privacy-filter model and opf training tools
• NVIDIA's nvidia/Nemotron-PII
• Gretel's gretelai/gretel-pii-masking-en-v1
• AI4Privacy's ai4privacy/pii-masking-openpii-1m

Citation

@misc{openmed_privacy_filter_nemotron_v2_2026,
  author       = {OpenMed},
  title        = {{OpenMed/privacy-filter-nemotron-v2}: second-generation Nemotron-schema privacy filter},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/OpenMed/privacy-filter-nemotron-v2}}
}