Model Overview

OPTML-Group/SimNPO-WMDP-zephyr-7b-beta is a 7-billion parameter language model derived from the HuggingFaceH4/zephyr-7b-beta base model. Its key differentiator is the application of the SimNPO (Simplicity Prevails: Rethinking Negative Preference Optimization) unlearning algorithm to remove specific knowledge related to the WMDP dataset.

Key Capabilities & Features

• Targeted Unlearning: Utilizes the novel SimNPO algorithm for efficient and effective removal of undesirable information from the base model.
• Research Focus: Primarily intended for research into LLM unlearning, offering a practical example of applying SimNPO.
• Performance Metrics: Evaluation results show that SimNPO achieves higher unlearning efficacy (e.g., 0.584 on 1-AccBio and 0.678 on 1-AccCyber) compared to the original model and other unlearning methods like NPO, while minimally impacting general capabilities (MMLU score of 0.471).
• Reproducible Methodology: The unlearning process and code base are available via github.com/OPTML-Group/Unlearn-Simple.

When to Use This Model

• LLM Unlearning Research: Ideal for researchers and developers exploring methods for removing specific data or behaviors from large language models.
• Privacy-Preserving AI: Useful for experimenting with techniques to enhance data privacy and compliance in deployed LLMs.
• Comparative Studies: Can be used as a baseline or comparison point for evaluating new unlearning algorithms against SimNPO.