gabrielloiseau/BERT-base-privacy

Distilling Human-Aligned Privacy Sensitivity Assessment from Large Language Models

This model is a lightweight encoder (150M parameters) based on BERT, designed for assessing the privacy sensitivity of textual data. It was introduced in the paper Distilling Human-Aligned Privacy Sensitivity Assessment from Large Language Models.

The model was distilled from Mistral Large 3 (675B) using a large-scale dataset of privacy-annotated texts across 10 diverse domains. It aims to preserve strong agreement with human judgments of privacy while being efficient enough for large-scale real-world deployment.

Model Details

• Task: Text Classification (Privacy Sensitivity Score 1-5)
• Model Type: BERT-base
• Distilled from: Mistral Large 3
• Repository: https://github.com/gabrielloiseau/privacy-distillation
• Paper: arXiv:2603.29497

How to Get Started with the Model

You can use the model directly with a Hugging Face pipeline:

from transformers import pipeline

classifier = pipeline("text-classification", model="gabrielloiseau/BERT-base-privacy")
result = classifier("Happy First Day of Spring!")
print(result)  # [{'label': '1', 'score': 0.98}]

The labels 1 through 5 represent the degree of privacy sensitivity, where 1 is the lowest and 5 is the highest.

Citation

@misc{loiseau2026distilling,
      title={Distilling Human-Aligned Privacy Sensitivity Assessment from Large Language Models}, 
      author={Gabriel Loiseau and Damien Sileo and Damien Riquet and Maxime Meyer and Marc Tommasi},
      year={2026},
      eprint={2603.29497},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2603.29497}, 
}