OpenAI Privacy Filter

OpenAI Privacy Filter

Finds and masks personal information such as names and emails in text in a single pass, and runs on your own machine. For teams that need to clean large amounts of text on-premises.

Made by openai 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: OpenAI Privacy Filter is a model that finds personally identifiable information in text and marks it so it can be masked. It reads a whole input at once and labels each token, then decodes clean spans from those labels.

It is built for data sanitization jobs where a team wants a model it can run on its own hardware, with settings it can tune.

Who it suits: It suits developers and data teams who must strip names, emails and similar details from text before they store, share or train on it.

Industry | Who uses it | How they use it | Example
Software | Developers | Add a masking step to a text pipeline | Hide names in logs before sharing
Data teams | Data engineers | Clean large text sets before training | Remove emails from a corpus
Support | Operations teams | Redact personal details in tickets | Mask a name in a customer message
Research | ML engineers | Fine-tune the filter on their own data | Adapt it to a new text style

If your text never holds personal details, or you need a filter for images or audio, you can skip it.

Where to use it: The model works in Python with the transformers library and in JavaScript with transformers.js, where the card shows it running with WebGPU.

• Python with transformers on Windows, macOS or Linux
• A web browser or Node app through transformers.js
• On-premises servers, since you run it yourself
• A laptop, given its small size

Use case | Who | What they get back
PII detection | Developers | Labeled spans such as private_person and private_email
Text masking | Data teams | Text with personal details marked for removal
On-premises cleaning | Security teams | Sanitized data that never leaves your servers

How to set it up: You need Python with the transformers library and PyTorch, or a JavaScript project with transformers.js. The weights come to about 2.8 GB in BF16.

• Python with pip, plus transformers and PyTorch
• Or Node with the Hugging Face transformers.js package
• About 2.8 GB of disk space for BF16 weights

1. Install the Hugging Face hub package.
2. Download the model files, or let the library fetch them by model id.
3. Load the model with the pipeline call.

pip install -U huggingface_hub
hf download openai/privacy-filter

How to use it: You pass text to a token-classification pipeline and read the labeled spans it returns.

from transformers import pipeline

classifier = pipeline(
    task="token-classification",
    model="openai/privacy-filter",
)
classifier("My name is Alice Smith")

Tool | What it does
pipeline | Runs token classification in a single call
AutoModelForTokenClassification | Gives direct access to the logits and labels
AutoTokenizer | Turns your text into model inputs

Example: You send the sentence "My name is Alice Smith" to the classifier. It labels the name as a private person, so you can replace that span with a mask.

For newcomers and experts: If you are new, run the short Python snippet above on one sentence with a name in it and read the labels that come back. That shows what the filter detects.

If you know the field, it is an autoregressive checkpoint converted into a bidirectional banded-attention token classifier, with a sparse mixture-of-experts feed-forward stack. Inference uses constrained Viterbi decoding to produce BIOES span labels.

Text is tokenized
One forward pass labels every token
Viterbi decoding builds spans
Spans are masked or removed

Good to know: The card lists a 128,000-token context window, so long text needs no chunking. It also describes preset operating points that trade precision against recall.

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the text you pass in and the model files you download, with no web service named

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: 102,911
• Likes: 1,831
• License: apache-2.0
• Quantized versions: 10
• Fine-tunes: 54
• Tags: transformers, onnx, safetensors, openai_privacy_filter, token-classification, transformers.js

Links

• Model: https://huggingface.co/openai/privacy-filter
• Files: https://huggingface.co/openai/privacy-filter/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download openai/privacy-filter

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