JoaoBoer/tofu_Llama-3.2-1B-Instruct_forget05_IdkNLL
The JoaoBoer/tofu_Llama-3.2-1B-Instruct_forget05_IdkNLL is a 1 billion parameter instruction-tuned Llama-3.2 model, specifically unlearned on the TOFU 'forget05' split using the IdkNLL method within the open-unlearning framework. This model serves as a weight-unlearning baseline and draft model for the Speculative-Decoding-Unlearning project. Its primary differentiator is its focus on demonstrating and evaluating machine unlearning capabilities, particularly in forgetting specific data while retaining general utility.
Loading preview...
Model Overview
This model, tofu_Llama-3.2-1B-Instruct_forget05_IdkNLL, is a 1 billion parameter instruction-tuned variant of the Llama-3.2 architecture. It has undergone a specific unlearning process on the TOFU forget05 dataset split, utilizing the IdkNLL method. This unlearning was performed within the open-unlearning framework.
Key Characteristics
- Unlearning Focus: The model's primary purpose is to serve as a baseline for evaluating machine unlearning, specifically for the Speculative-Decoding-Unlearning project.
- Methodology: It employs the IdkNLL method for weight-unlearning, with specific hyperparameters such as
gamma: -1.0,alpha: 2, andretain_loss_type: NLL. - Evaluation Metrics: Performance is assessed using TOFU summary metrics, including
exact_memorization(0.7499),extraction_strength(0.1529), andforget_Q_A_PARA_Prob(0.0887), indicating its ability to forget specific information.
When to Use This Model
This model is particularly useful for:
- Research in Machine Unlearning: Ideal for researchers and developers exploring techniques for making models forget specific training data.
- Benchmarking Unlearning Methods: It provides a concrete baseline for comparing new unlearning algorithms against a known, established method (IdkNLL).
- Understanding Forgetting Mechanisms: Useful for analyzing how different unlearning parameters impact a model's ability to forget while attempting to preserve general utility.