This is a BERT Large model fine-tuned to detect privacy violations in text, such as sharing of personally identifiable information (PII) or sensitive data. It is trained on a dataset of labeled examples of privacy violations and non-violations.
Performance
This large model achieves the following performance metrics on a held-out test set:
Metric
Value
Accuracy
0.9584
F1 Score
0.9569
Precision
0.9621
Recall
0.9517
These metrics indicate that the model is effective at identifying privacy violations while minimizing false positives.
Limitations
The model was trained on a dataset of nearly 1 million examples in varying topics and styles, but may not generalize to all contexts
It limited to English text
This current iteration used a dataset where each example is between 20 and 120 words in length, so performance on much longer texts is untested (e.g. full documents may require chunking)
The model may not detect all types of privacy violations, especially if they are subtle or context-dependent
How to Use
You can use this model for text classification tasks related to privacy moderation. Here's an example of how to use it with the Hugging Face Transformers library:
python
1from transformers import AutoModelForSequenceClassification, AutoTokenizer
2import torch
3import numpy as np
4import pandas as pd
56# Load the model and tokenizer7model_name ="PL-RnD/privacy-moderation-large"8tokenizer = AutoTokenizer.from_pretrained(model_name)9model = AutoModelForSequenceClassification.from_pretrained(model_name)10# Example text11texts =[12"Here is my credit card number: 1234-5678-9012-3456",13"This is a regular message without sensitive information.",14"For homeowners insurance, select deductibles from $500 to $2,500. Higher deductibles lower premiums.",15"Solidarity: My enrollment includes my kid's braces at $4,000 total—family strained. Push for orthodontic expansions. Email blast to reps starting now.",16]17# Tokenize the input18inputs = tokenizer(texts, return_tensors="pt", padding=True, truncation=True)19# Get model predictions20with torch.no_grad():21 outputs = model(**inputs)2223logits = outputs.logits
24predictions = torch.argmax(logits, dim=-1)25# Convert predictions to labels26labels =["non-violation","violation"]27predicted_labels =[labels[pred]for pred in predictions.numpy()]28# Display results29df = pd.DataFrame({"text": texts,"label": predicted_labels})30print(df)
This will output a DataFrame with the original texts and their predicted labels (either "violation" or "non-violation"). Example output:
text label
0 Here is my credit card number: 1234-5678-9012-... violation
1 This is a regular message without sensitive in... non-violation
2 For homeowners insurance, select deductibles f... non-violation
3 Solidarity: My enrollment includes my kid's br... violation
Intended Use
This model is intended to flag privacy concerns that a privacy conscious person would expect to keep private, such as: addresses, phone numbers, e-mails, passwords, health details, relationship drama, financial numbers, political opinions, or sexual preferences.
The motivating use-case for this model is to reside client-side (or in a trusted/internal environment) to review user-generated text content before it is sent to a server or third-party service, in order to prevent accidental sharing of sensitive information. For example:
Filter and act as an A:B router for public vs private LLMS (i.e. like using this with Pipelines in Open-WebUI). If the text is flagged as a privacy violation, it can be routed to a local/private LLM instance instead of a public one.
Block or warn users when they attempt to share sensitive information in chat applications
Load the model in a browser using libraries like ONNX.js or TensorFlow.js to perform client-side moderation
Ultimately, arguing that you don't care about the right to privacy because you have nothing to hide is no different than saying you don't care about free speech because you have nothing to say. - Edward Snowden
