JoaoBoer/tofu_Llama-3.2-1B-Instruct_forget01_GradDiff
JoaoBoer/tofu_Llama-3.2-1B-Instruct_forget01_GradDiff is a 1 billion parameter instruction-tuned Llama-3.2 model, specifically unlearned on the TOFU 'forget01' split using the GradDiff method. Developed within the open-unlearning framework, this model serves as a weight-unlearning baseline for the Speculative-Decoding-Unlearning project. It demonstrates targeted forgetting capabilities while maintaining a model utility of 0.5883 and an exact memorization score of 0.9217.
Loading preview...
Model Overview
JoaoBoer/tofu_Llama-3.2-1B-Instruct_forget01_GradDiff is a 1 billion parameter instruction-tuned Llama-3.2 model that has undergone a specific unlearning process. It was unlearned on the TOFU forget01 dataset split utilizing the GradDiff method, implemented within the open-unlearning framework. This model is primarily intended as a weight-unlearning baseline or draft model for the Speculative-Decoding-Unlearning research project.
Key Characteristics
- Unlearning Method: Employs the GradDiff method for targeted forgetting.
- Dataset: Unlearned on the
forget01split of the TOFU dataset. - Framework: Developed using the open-unlearning framework.
- Hyperparameters: Configured with
gamma: 1.0,alpha: 5, andretain_loss_type: NLL.
Performance Metrics (TOFU Evaluation)
This model exhibits specific unlearning characteristics as measured by TOFU summary metrics:
- Exact Memorization: 0.9217
- Extraction Strength: 0.3464
- Forget Q&A PARA Prob: 0.0655
- Forget Quality: 0.0068
- Model Utility: 0.5883
- MIA Loss: 0.9800
These metrics indicate its effectiveness in forgetting specific information while attempting to preserve overall utility.