Machine Unlearning on TOFU finetuned Llama-2-7b-chat (forget set)
99.51ProbabilityBefore
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| BeforeStatus=Pre-unlearning, Model=Llama-2-7b-chat2026.04 | 99.51 | 94.93 | 53.06 | 0.13 | 71.33 | 16.67 | 5.08 | 62.77 | |
| BeforeBackbone=Llama-2-7b-chat2026.04 | 99.51 | 94.93 | 53.06 | 0.13 | 71.33 | 16.67 | 5.08 | 62.77 | |
| DirectQABaseline Setting=GA+GD, Backbone=Llama-2-7b-chat2026.04 | 98.43 | 91.98 | 53.95 | 0.68 | 70.15 | 18.52 | 6.16 | 63.33 | |
| TOFUBaseline Setting=GA+KL, Backbone=Llama-2-7b-chat2026.04 | 47.09 | 44.09 | 56.44 | 5.41 | 59.23 | 16.17 | 5.16 | 59.43 | |
| NPOGeneration Method=DirectQA, Model=Llama-2-7b-chat2026.04 | 43.93 | 49.33 | 56.37 | 5.41 | 52.04 | 8.17 | 2.2 | 57.43 | |
| GAGeneration Method=DirectQA, Model=Llama-2-7b-chat2026.04 | 42.6 | 45.21 | 50.29 | 2.86 | 53.78 | 8.3 | 2.12 | 57.83 | |
| DirectQABaseline Setting=GA+KL, Backbone=Llama-2-7b-chat2026.04 | 29.16 | 40.92 | 49.5 | 5.41 | 50.69 | 9.2 | 2.72 | 56.88 | |
| MAGEBaseline Setting=GA+GD, Backbone=Llama-2-7b-chat2026.04 | 16.08 | 44.16 | 52.83 | 16.5 | 47.48 | 1.43 | 0.48 | 54.36 | |
| MAGEBaseline Setting=GA+KL, Backbone=Llama-2-7b-chat2026.04 | 15.8 | 33.59 | 47.16 | 9.71 | 47.67 | 0.57 | 0.28 | 53.62 | |
| GAGeneration Method=MAGE, Model=Llama-2-7b-chat2026.04 | 14.92 | 45 | 52.85 | 5.41 | 44.49 | 0.46 | 0.24 | 51.55 | |
| TOFUBaseline Setting=GA+GD, Backbone=Llama-2-7b-chat2026.04 | 10.31 | 16.31 | 73.65 | 26.57 | 18.34 | 52.51 | 39.04 | 34.97 | |
| NPOGeneration Method=MAGE, Model=Llama-2-7b-chat2026.04 | 7.32 | 24.87 | 56.7 | 9.71 | 26.13 | 3.84 | 0.46 | 39.49 | |
| NPOGeneration Method=TOFU, Model=Llama-2-7b-chat2026.04 | 3.59 | 3.38 | 62.72 | 16.5 | 16.38 | 24.02 | 9.48 | 29.51 | |
| GAGeneration Method=TOFU, Model=Llama-2-7b-chat2026.04 | 3.53 | 7.14 | 59.03 | 16.5 | 19.06 | 19.22 | 7.25 | 33.09 |