privacy-filter-nemotron-v2

privacy-filter-nemotron-v2

Finds and masks personal data in text, across 55 fine-grained categories such as names, contact details and IDs. It suits teams that redact text before sharing or storing it.

Made by OpenMed on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: privacy-filter-nemotron-v2 is a token classifier that spots personal data in text so you can mask it. It is the second generation of the OpenMed Nemotron-schema checkpoints. It labels 55 kinds of personal data, from names and contact details to account numbers. The card says v2 masks more reliably than the original.

Who it suits: It suits teams that must remove personal data from text before sharing, storing or analyzing it. You need some Python to run it.

Industry | Who uses it | How they use it | Example
Healthcare | Data and compliance teams | Mask patient identifiers in notes before they leave the system | Hiding a name, phone number and record number in a note
Customer support | Support operations | Redact personal details from transcripts and logs | Cleaning a chat log before analysis
Legal and compliance | Privacy officers | Strip identifiers from documents before review | Redacting a contract excerpt before sharing it
Software | Data engineers | Add a redaction step to a text pipeline | Masking text before it goes to another service

People who need a single dataset lineage can use the original Nemotron checkpoint instead.

Where to use it:

• Python with the OpenMed package
• The transformers library, as a token classification pipeline
• The opf command line tool from OpenAI's privacy filter project
• Windows, macOS or Linux

Use case | Who | What they get back
Finding personal data | Compliance teams | Each detected item with its label and text
Masking text | Data engineers | The same text with personal data masked
General redaction | Privacy officers | Typed labels you can use to choose what to hide

How to set it up:

• Python on your computer
• The OpenMed package with its Hugging Face extra, or the transformers library
• About 2.8 GB of disk for the BF16 weights

1. Install the Hugging Face hub package.
2. Download the model files.
3. Install the OpenMed package, or load the model with transformers.

pip install -U huggingface_hub

hf download OpenMed/privacy-filter-nemotron-v2

pip install -U "openmed[hf]"

How to use it: With OpenMed, call extract_pii to list the personal data in a text, and call deidentify with the mask method to hide it. For best production results, the card advises BIOES-aware decoding and merging overlapping or adjacent spans before masking.

from openmed import extract_pii, deidentify

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

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)

Tool | What it does
extract_pii | Lists the personal data found in a text, with a label for each
deidentify | Returns the text with the personal data masked
opf redact | Runs redaction from the command line with a chosen checkpoint

Example: You pass a sentence that holds a patient's name, birth date, record number, phone number and email. extract_pii prints each item with its label. deidentify with the mask method returns the sentence with those items hidden.

For newcomers and experts: A token classifier tags each piece of text with a label, and here the labels mark personal data. A first small step is to run the OpenMed sample on one made-up sentence and read the labels.

Its base is openai/privacy-filter, a mixture-of-experts classifier of 1.4B parameters, about 50M of which are active for each token. It predicts 221 BIOES classes, made of one outside class and four tags for each of the 55 categories. Training was a full fine-tune with a broader mix and a recall-oriented recipe.

Text with personal data
Model tags each token with a label
Spans are merged and decoded
Personal data is masked

Good to know:

• Languages: the record lists several European languages, including Bulgarian, Czech, Danish, German, Greek, English, Spanish and Estonian.
• License: Apache 2.0.
• Cost to run: none named. The steps need no API key or paid service.
• Data it reaches: it downloads files from the Hugging Face Hub, then processes the text you give it. The transformers sample turns on trust_remote_code, so it runs code from the model repository.
• Not for: skipping human review. The card text provided gives no accuracy figures.

On Hugging Face, 13 November 2026

• Task: token classification
• Library: transformers
• Parameters: 1.4B
• Weights: about 2.8 GB in BF16; about 700 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 154,398
• Likes: 57
• License: apache-2.0
• Languages: bg, cs, da, de, el, en, es, et
• Based on: openai/privacy-filter
• Quantized versions: 3
• Fine-tunes: 2
• Tags: transformers, safetensors, openai_privacy_filter, token-classification, privacy, redaction, nemotron, openmed, openai-privacy-filter

Links

• Model: https://huggingface.co/OpenMed/privacy-filter-nemotron-v2
• Files: https://huggingface.co/OpenMed/privacy-filter-nemotron-v2/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download OpenMed/privacy-filter-nemotron-v2

Then follow the usage in its model card. Nothing in this kit runs by itself.