tofu_Llama-3.1-8B-Instruct_forget05_PDU
open-unlearning/tofu_Llama-3.1-8B-Instruct_full unlearned on the TOFU forget05 split with PDU, 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: 100
retain_loss_type: NLL
retain_loss_eps: 0.3
primal_dual: True
dual_step_size: 5
dual_update_upon: step
dual_warmup_epochs: 5
loss_names: ['forget_loss', 'retain_loss']
TOFU summary metrics
| metric | value |
|---|---|
| exact_memorization | 0.0170 |
| extraction_strength | 0.0327 |
| forget_Q_A_PARA_Prob | 0.0000 |
| forget_Q_A_gibberish | 0.0767 |
| forget_quality | 0.0000 |
| forget_truth_ratio | 0.9337 |
| mia_loss | 0.0001 |
| mia_min_k | 0.0000 |
| mia_min_k_plus_plus | 0.0152 |
| mia_zlib | 0.0294 |
| model_utility | 0.6685 |
| privleak | 55.3277 |