tofu_Llama-3.1-8B-Instruct_forget10_RMU
open-unlearning/tofu_Llama-3.1-8B-Instruct_full unlearned on the TOFU forget10 split with RMU, 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: 1
retain_loss_type: EMBED_DIFF
steering_coeff: 1
module_regex: model\.layers\.5
trainable_params_regex: ['.*']
TOFU summary metrics
| metric | value |
|---|---|
| exact_memorization | 0.1273 |
| extraction_strength | 0.0326 |
| forget_Q_A_PARA_Prob | 0.0014 |
| forget_Q_A_gibberish | 0.2818 |
| forget_quality | 0.0000 |
| forget_truth_ratio | 0.7325 |
| mia_loss | 0.0189 |
| mia_min_k | 0.0229 |
| mia_min_k_plus_plus | 0.9275 |
| mia_zlib | 0.0082 |
| model_utility | 0.6750 |
| privleak | 57.6383 |