tofu_Llama-3.2-3B-Instruct_forget05_SatImp
open-unlearning/tofu_Llama-3.2-3B-Instruct_full unlearned on the TOFU forget05 split with SatImp, trained with the open-unlearning framework. Used as a weight-unlearning baseline / draft model in the Speculative-Decoding-Unlearning project.
Full training config: .hydra/config.yaml. TOFU evaluation outputs: evals/.
Method hyperparameters
gamma: 1.0
alpha: 0.1
retain_loss_type: NLL
beta1: 5.0
beta2: 1.0
TOFU summary metrics
| metric | value |
|---|---|
| exact_memorization | 0.3267 |
| extraction_strength | 0.0414 |
| forget_Q_A_PARA_Prob | 0.0103 |
| forget_Q_A_gibberish | 0.4607 |
| forget_quality | 0.0000 |
| forget_truth_ratio | 0.7660 |
| mia_loss | 0.0341 |
| mia_min_k | 0.0991 |
| mia_min_k_plus_plus | 0.9885 |
| mia_zlib | 0.0242 |
| model_utility | 0.6324 |
| privleak | 40.8407 |