OPTML-Group/SimNPO-TOFU-forget05-Llama-2-7b-chat

Model Overview

This model, SimNPO-TOFU-forget05-Llama-2-7b-chat, is a 7 billion parameter Llama-2-chat based language model developed by OPTML-Group. Its primary distinction lies in its application of the SimNPO unlearning algorithm to specifically remove information related to the TOFU - Forget05 dataset. This unlearning process aims to demonstrate effective data removal from large language models.

Key Capabilities

• Targeted Unlearning: Utilizes the SimNPO algorithm to selectively forget specific data points, achieving a Forgetting Quality (FQ) of 0.99, comparable to a full retraining (Retrain FQ: 1.00).
• Utility Preservation: While unlearning, the model largely preserves its general utility, showing a Model Utility (MU) of 0.58, close to the original model's 0.62.
• Research into LLM Unlearning: Serves as a practical example and benchmark for the effectiveness of the SimNPO method, as detailed in the research paper "Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning".

When to Use This Model

This model is particularly useful for:

• Researchers studying machine unlearning and data privacy in large language models.
• Evaluating the effectiveness of different unlearning algorithms against a known baseline.
• Demonstrating the ability to remove specific information from a pre-trained LLM while minimizing impact on general performance.