OPTML-Group/NPO-GP-WMDP
The OPTML-Group/NPO-GP-WMDP is a 7 billion parameter causal language model derived from HuggingFaceH4/zephyr-7b-beta, specifically fine-tuned for unlearning tasks. It utilizes the NPO method with Gradient Penalty (GP) for smoothness optimization, focusing on the wmdp-bio dataset. This model is designed for research into unlearning capabilities in large language models, particularly in making them resilient to relearning attacks.
Loading preview...
Model Overview
OPTML-Group/NPO-GP-WMDP is a 7 billion parameter language model built upon the HuggingFaceH4/zephyr-7b-beta architecture. Its primary focus is on unlearning, a critical area in AI safety and data privacy. This model implements the NPO (Neural Perceptron Optimization) method, enhanced with Gradient Penalty (GP) for improved smoothness during the unlearning process.
Key Capabilities
- Targeted Unlearning: Specifically trained to unlearn information related to the
wmdp-biotask from the 🤗datasets/cais/wmdp dataset. - Smoothness Optimization: Incorporates Gradient Penalty (GP) to ensure a more stable and effective unlearning process.
- Research Focus: Developed as part of research into making LLM unlearning resilient to relearning attacks, as detailed in the paper "Towards LLM Unlearning Resilient to Relearning Attacks: A Sharpness-Aware Minimization Perspective and Beyond" (arXiv:2502.05374).
When to Use This Model
This model is ideal for researchers and developers interested in:
- Exploring and evaluating LLM unlearning techniques.
- Investigating methods to enhance unlearning resilience against relearning attacks.
- Working with models specifically designed for data removal or privacy-preserving AI.
It provides a practical implementation of NPO with GP for unlearning, offering a valuable tool for advancing research in this specialized domain.