JoaoBoer/tofu_Llama-3.2-3B-Instruct_forget01_RMU
The JoaoBoer/tofu_Llama-3.2-3B-Instruct_forget01_RMU is a 3.2 billion parameter instruction-tuned Llama-3.2 model, developed by JoaoBoer, that has undergone unlearning on the TOFU 'forget01' split using the RMU method. It is specifically designed as a weight-unlearning baseline for the Speculative-Decoding-Unlearning project. This model demonstrates specific metrics related to forgetting and retention, making it suitable for research in machine unlearning and privacy-preserving AI.
Loading preview...
Model Overview
The JoaoBoer/tofu_Llama-3.2-3B-Instruct_forget01_RMU is a 3.2 billion parameter instruction-tuned model based on the Llama-3.2 architecture. It was developed by JoaoBoer as part of the Speculative-Decoding-Unlearning project.
Key Characteristics
This model is a result of applying machine unlearning techniques to an existing open-unlearning/tofu_Llama-3.2-3B-Instruct_full model. Specifically, it was unlearned on the TOFU forget01 split using the RMU (Retain-Memory-Unlearning) method, implemented within the open-unlearning framework.
Unlearning Performance Metrics
The model's unlearning effectiveness is quantified by several TOFU summary metrics, including:
- exact_memorization: 0.6484
- extraction_strength: 0.0512
- forget_Q_A_PARA_Prob: 0.0442
- forget_quality: 0.7659
- model_utility: 0.6446
- privleak: -28.3898
These metrics indicate its performance in forgetting specific information while attempting to retain general utility.
Use Cases
This model is primarily intended for:
- Research in machine unlearning: Serving as a baseline for weight-unlearning experiments.
- Privacy-preserving AI studies: Investigating methods to remove specific data from trained models.
- Development of unlearning techniques: Exploring the impact of different unlearning hyperparameters and strategies.