OPTML-Group/NPO-WA-WMDP
OPTML-Group's NPO-WA-WMDP is a 7 billion parameter causal language model derived from HuggingFaceH4/zephyr-7b-beta, specifically unlearned using the NPO method with Weight Averaging (WA) for the WMDP-bio task. This model focuses on demonstrating unlearning capabilities, particularly in making LLMs resilient to relearning attacks. It is designed for research into model unlearning and privacy-preserving AI, offering insights into techniques for removing specific information from trained models.
Loading preview...
Model Overview
OPTML-Group/NPO-WA-WMDP is a 7 billion parameter language model based on the HuggingFaceH4/zephyr-7b-beta architecture. Its primary distinction lies in its application of unlearning techniques, specifically the NPO (Neural Pruning Optimization) method combined with Weight Averaging (WA). This model was unlearned on the WMDP-bio dataset from cais/wmdp.
Key Capabilities & Focus
- Model Unlearning: Demonstrates the application of NPO and Weight Averaging for removing specific information (the WMDP-bio task) from a pre-trained LLM.
- Research into Unlearning Resilience: Developed as part of research exploring LLM unlearning that is resilient to relearning attacks, as detailed in the associated paper "Towards LLM Unlearning Resilient to Relearning Attacks: A Sharpness-Aware Minimization Perspective and Beyond" (arXiv:2502.05374).
- Privacy-Preserving AI: Contributes to the field of privacy-preserving machine learning by providing a practical example of unlearning.
Use Cases
This model is particularly suited for:
- Academic Research: Investigating and developing new unlearning algorithms for large language models.
- Evaluating Unlearning Effectiveness: Testing the robustness and efficiency of unlearning methods against various attacks.
- Understanding Model Behavior: Analyzing how unlearning impacts model performance and knowledge retention on specific tasks.
Developers interested in the practical implementation of unlearning techniques and their theoretical underpinnings will find this model and its associated codebase (github.com/OPTML-Group/Unlearn-Smooth) valuable.