tofu_Llama-3.1-8B-Instruct_forget01_GradDiff
open-unlearning/tofu_Llama-3.1-8B-Instruct_full unlearned on the TOFU forget01 split with GradDiff, 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: 5
retain_loss_type: NLL
TOFU summary metrics
| metric | value |
|---|---|
| exact_memorization | 0.9605 |
| extraction_strength | 0.5235 |
| forget_Q_A_PARA_Prob | 0.0434 |
| forget_Q_A_gibberish | 0.8753 |
| forget_quality | 0.0286 |
| forget_truth_ratio | 0.5078 |
| mia_loss | 0.9731 |
| mia_min_k | 0.9725 |
| mia_min_k_plus_plus | 0.8813 |
| mia_zlib | 0.9806 |
| model_utility | 0.6313 |
| privleak | -94.5000 |