JoaoBoer/tofu_Llama-3.1-8B-Instruct_forget01_UNDIAL

Model Overview

This model, tofu_Llama-3.1-8B-Instruct_forget01_UNDIAL, is an 8 billion parameter instruction-tuned variant of Llama-3.1. Its primary distinction lies in its application of unlearning techniques. Specifically, it has been unlearned on the TOFU forget01 dataset split using the UNDIAL method, developed within the open-unlearning framework.

Key Characteristics

• Unlearning Focus: Designed to evaluate and demonstrate the removal of specific information (TOFU forget01 split) from the model's knowledge base.
• Methodology: Utilizes the UNDIAL unlearning method, serving as a weight-unlearning baseline for the Speculative-Decoding-Unlearning project.
• Llama-3.1 Base: Built upon the Llama-3.1-8B-Instruct architecture, retaining its general capabilities prior to unlearning.
• Context Length: Supports a substantial context window of 32768 tokens.

Unlearning Performance (TOFU Metrics)

The model's unlearning effectiveness is quantified by several TOFU summary metrics, indicating its ability to forget specific data while attempting to preserve overall utility:

• exact_memorization: 0.4404
• forget_Q_A_PARA_Prob: 0.0845
• forget_quality: 0.1650
• model_utility: 0.4572

Use Cases

This model is particularly relevant for research and development in:

• Machine Unlearning: Studying the efficacy of unlearning algorithms and their impact on model performance.
• Privacy-Preserving AI: Exploring methods to remove sensitive or unwanted information from trained models.
• Model Editing: Investigating techniques for modifying model behavior post-training without full retraining.