JoaoBoer/tofu_Llama-3.2-1B-Instruct_forget10_IdkNLL
The JoaoBoer/tofu_Llama-3.2-1B-Instruct_forget10_IdkNLL is a 1 billion parameter Llama-3.2-Instruct model that has undergone unlearning on the TOFU 'forget10' split using the IdkNLL method. Developed within the open-unlearning framework, this model serves as a weight-unlearning baseline for the Speculative-Decoding-Unlearning project. It is specifically designed to evaluate and demonstrate unlearning capabilities, focusing on privacy and memorization metrics.
Loading preview...
Model Overview
The JoaoBoer/tofu_Llama-3.2-1B-Instruct_forget10_IdkNLL is a 1 billion parameter instruction-tuned Llama-3.2 model that has been specifically modified through an unlearning process. This model was unlearned on the TOFU forget10 dataset split utilizing the IdkNLL method, implemented within the open-unlearning framework. It functions as a baseline model for weight-unlearning experiments within the Speculative-Decoding-Unlearning project.
Key Characteristics
- Unlearning Focus: Designed to demonstrate and evaluate the effectiveness of unlearning specific data (the
forget10split of TOFU). - Methodology: Employs the IdkNLL method for unlearning, with specific hyperparameters like
gamma: -1.0andalpha: 2. - Evaluation Metrics: Provides detailed metrics on unlearning performance, including
exact_memorization(0.7590),extraction_strength(0.1415),forget_Q_A_PARA_Prob(0.0945), andprivleak(-91.6290), indicating its ability to forget while retaining utility.
Use Cases
This model is primarily intended for:
- Research in Machine Unlearning: Ideal for researchers studying methods to remove specific information from trained language models.
- Privacy-Preserving AI: Useful for exploring techniques to enhance data privacy in LLMs by selectively forgetting data.
- Baseline Comparisons: Serves as a comparative baseline for new unlearning algorithms and frameworks.