GPT-PDVS1-Low

GPT-PDVS1-Low is an experimental open-source text-generating AI designed for testing vulnerabilities in GPT-type models relating to the gathering, retention, and possible later dissemination (whether in accurate or distorted form) of individuals’ personal data.

GPT-PDVS1-Low is the member of the larger “GPT Personal Data Vulnerability Simulator” (GPT-PDVS) model family that has been fine-tuned on a text corpus to which 200 of its 18,000 paragraphs (or roughly 1.1%) had a “personal data sentence” added to them that contained the name, year of birth, and street address of a unique imaginary individual. Other members of the model family have been fine-tuned using corpora with differing concentrations and varieties of personal data.

Model description

The model is a fine-tuned version of GPT-2 that has been trained on a text corpus containing 18,000 paragraphs from pages in the English-language version of Wikipedia that has been adapted from the “Quoref (Q&A for Coreference Resolution)” dataset available on Kaggle.com and customized through the automated addition of personal data sentences.

Intended uses & limitations

This model has been designed for experimental research purposes; it isn’t intended for use in a production setting or in any sensitive or potentially hazardous contexts.

Training procedure and hyperparameters

The model was fine-tuned using a Tesla T4 with 16GB of GPU memory. The following hyperparameters were used during training:

• optimizer: {'name': 'AdamWeightDecay', 'learning_rate': {'class_name': 'ExponentialDecay', 'config': {'initial_learning_rate': 0.0005, 'decay_steps': 500, 'decay_rate': 0.95, 'staircase': False, 'name': None}}, 'decay': 0.0, 'beta_1': 0.9, 'beta_2': 0.999, 'epsilon': 1e-07, 'amsgrad': False, 'weight_decay_rate': 0.01}
• training_precision: float32
• epochs: 8

Framework versions

• Transformers 4.27.1
• TensorFlow 2.11.0
• Datasets 2.10.1
• Tokenizers 0.13.2