OpenAI’s Open‑Source Privacy Filter: Local PII Detection Without Server Upload (Apache 2.0)

Model Overview

OpenAI released an open‑source Privacy Filter model under Apache 2.0. The model has 1.5 B parameters, with a sparse Mixture‑of‑Experts architecture that activates only 50 M parameters per inference, enabling execution on ordinary hardware. It detects eight predefined personally identifiable information (PII) categories—names, addresses, phone numbers, email addresses, account identifiers, secret data, etc.—and supports multilingual input. The model runs fully locally via Transformers.js and WebGPU, so inference stays on the user’s device and no text is transmitted to a server.

Key technical specifications:

Context window: 128 k tokens, allowing processing of long documents without chunking.

Runtime controls to trade precision against recall.

Fine‑tuning capability to adapt the eight categories to domain‑specific rules.

Usage Examples

Python (Transformers) – quick pipeline call :

# Use pipeline for quick inference
from transformers import pipeline

pipe = pipeline("token-classification", model="openai/privacy-filter")

# Direct model loading
from transformers import AutoModel
model = AutoModel.from_pretrained("openai/privacy-filter", dtype="auto")

Browser (Transformers.js) – WebGPU execution :

// Install dependency
npm i @huggingface/transformers
import { pipeline } from '@huggingface/transformers';

// Initialize pipeline on WebGPU
const classifier = await pipeline('token-classification', 'openai/privacy-filter', { device: 'webgpu', dtype: 'q4' });

// Example input
const input = 'My name is Harry Potter and my email is [email protected].';
const output = await classifier(input, { aggregation_strategy: 'simple' });
console.dir(output, { depth: null });

Sample output for the above input:

[
  {
    entity_group: 'private_person',
    score: 0.9999957978725433,
    word: ' Harry Potter'
  },
  {
    entity_group: 'private_email',
    score: 0.9999990728166368,
    word: ' [email protected]'
  }
]

Adoption Metrics

In the first month after release, the repository on Hugging Face recorded over 305 000 downloads, 36 fine‑tuned variants, and 38 related application spaces. Users reported very high detection accuracy for email addresses; Chinese‑language PII detection was noted as less reliable.

Official Limitations

Only the eight predefined PII types are supported; adding new categories requires fine‑tuning and cannot be changed at runtime.

Performance degrades on non‑English or non‑Latin scripts, with higher miss rates for low‑frequency names and identifier formats.

False positives may occur, such as labeling common nouns or public entities as private data, and some domain‑specific identifiers may be missed.

In high‑sensitivity scenarios, human review remains necessary; the model should be treated as an assistive tool rather than a fully automated compliance solution.

Model repository: https://huggingface.co/openai/privacy-filter

Demo space (WebGPU): https://huggingface.co/spaces/webml-community/privacy-filter-webgpu