tofu_Llama-3.1-8B-Instruct_forget01_SimNPO
open-unlearning/tofu_Llama-3.1-8B-Instruct_full unlearned on the TOFU forget01 split with SimNPO, 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: 0.125
alpha: 1
retain_loss_type: NLL
delta: 1
beta: 3.5
TOFU summary metrics
| metric | value |
|---|---|
| exact_memorization | 0.7239 |
| extraction_strength | 0.0563 |
| forget_Q_A_PARA_Prob | 0.0729 |
| forget_Q_A_gibberish | 0.9259 |
| forget_quality | 0.9188 |
| forget_truth_ratio | 0.6905 |
| mia_loss | 0.8072 |
| mia_min_k | 0.8000 |
| mia_min_k_plus_plus | 0.5394 |
| mia_zlib | 0.7187 |
| model_utility | 0.6254 |
| privleak | -60.0000 |