Knowledge Unlearning on TOFU forget01 (LLaDA-8B-Base)
35.55ForgetLPrandom_layer
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| random_layerBackbone=LLaDA-8B-Base, LoRA rank (r)=128, LoRA epochs (e)=102026.06 | 35.55 | 34.71 | 13.27 | 6.08 | 0.513 | |
| rmu_likeBackbone=LLaDA-8B-Base, LoRA rank (r)=128, LoRA epochs (e)=102026.06 | 35.58 | 34.86 | 12.95 | 5.95 | 0.607 | |
| no_editBackbone=LLaDA-8B-Base, LoRA rank (r)=128, LoRA epochs (e)=102026.06 | 35.7 | 34.76 | 13.39 | 6.11 | 0.514 | |
| TimeROMEalpha (α)=0.25, q=2, Backbone=LLaDA-8B-Base, LoRA rank (r)=128, LoRA epochs (e)=102026.06 | 36.12 | 35.09 | 13.61 | 6.64 | 0.534 | |
| TimeROMEalpha (α)=0.5, q=2, Backbone=LLaDA-8B-Base, LoRA rank (r)=128, LoRA epochs (e)=102026.06 | 37.55 | 36.33 | 14.34 | 7.47 | 0.578 | |
| GDConfig=safe lr, Backbone=LLaDA-8B-Base, LoRA rank (r)=128, LoRA epochs (e)=102026.06 | 37.86 | 34.65 | 13.07 | 6.04 | 0.524 | |
| GAConfig=safe lr, Backbone=LLaDA-8B-Base, LoRA rank (r)=128, LoRA epochs (e)=102026.06 | 38.21 | 34.94 | 13.41 | 6.32 | 0.511 | |
| MEMITBackbone=LLaDA-8B-Base, LoRA rank (r)=128, LoRA epochs (e)=102026.06 | 39.05 | 35.4 | 13.51 | 6.71 | 0.498 | |
| NPOBackbone=LLaDA-8B-Base, LoRA rank (r)=128, LoRA epochs (e)=102026.06 | 39.12 | 35.12 | 13.55 | 6.29 | 0.531 | |
| UnKEBackbone=LLaDA-8B-Base, LoRA rank (r)=128, LoRA epochs (e)=102026.06 | 41.27 | 36.18 | 14.02 | 7.18 | 0.547 | |
| TimeROMEalpha (α)=1.0, q=2, Backbone=LLaDA-8B-Base, LoRA rank (r)=128, LoRA epochs (e)=102026.06 | 42.04 | 40.27 | 16.79 | 9.56 | 0.562 | |
| SimNPOBackbone=LLaDA-8B-Base, LoRA rank (r)=128, LoRA epochs (e)=102026.06 | 42.84 | 36.71 | 14.18 | 7.02 | 0.557 | |
| Adaptive-RMUBackbone=LLaDA-8B-Base, LoRA rank (r)=128, LoRA epochs (e)=102026.06 | 45.31 | 38.42 | 15.24 | 7.91 | 0.578 | |
| AlphaEditBackbone=LLaDA-8B-Base, LoRA rank (r)=128, LoRA epochs (e)=102026.06 | 47.62 | 39.83 | 16.05 | 8.42 | 0.551 | |
| act_steerImplementation=RepE, Backbone=LLaDA-8B-Base, LoRA rank (r)=128, LoRA epochs (e)=102026.06 | 50.17 | 49.79 | 22.45 | 14.93 | 0.388 | |
| TimeROMEalpha (α)=2.0, q=4, Variant=BEST, Backbone=LLaDA-8B-Base, LoRA rank (r)=128, LoRA epochs (e)=102026.06 | 118.94 | 89.08 | 25.24 | 16.87 | 0.153 |