Machine Unlearning on TOFU (Forget05)
6.62RelevanceTRU
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| TRUModel=Zephyr-7B-beta2026.02 | 6.62 | 6.64 | 7.73 | 6.6 | 3.68 | 4.41 | — | |
| POModel=Zephyr-7B-beta2026.02 | 3.57 | 1.24 | 0.09 | 1.24 | 0.92 | 0.89 | — | |
| GradAscentModel=Zephyr-7B-beta2026.02 | 2.84 | 0.02 | 0 | 0 | 0 | 0 | — | |
| GradDiffModel=Zephyr-7B-beta2026.02 | 1.45 | 0 | 0 | 0.03 | 0.23 | 0.27 | — | |
| KLModel=Zephyr-7B-beta2026.02 | 1.37 | 0.05 | 0.02 | 1.6 | 2.17 | 2.94 | — | |
| RMUModel=Zephyr-7B-beta2026.02 | 0.17 | 0 | 0.03 | 5.71 | 6.82 | 7.49 | — | |
| WGAModel=Zephyr-7B-beta2026.02 | 0.13 | 0 | 0.04 | 5.37 | 6.7 | 7.08 | — | |
| NPOModel=Zephyr-7B-beta2026.02 | 0.1 | 0 | 0 | 3.74 | 5.67 | 6.64 | — | |
| Guardrail ClassifierBackbone=Llama-3-8B-Instruct2026.05 | — | — | — | — | — | — | 1 | |
| Guardrail ClassifierBackbone=Llama-2-7B-chat-hf2026.05 | — | — | — | — | — | — | 1 | |
| Guardrail FilterBackbone=Llama-3-8B-Instruct2026.05 | — | — | — | — | — | — | 0.495 | |
| Guardrail FilterBackbone=Llama-2-7B-chat-hf2026.05 | — | — | — | — | — | — | 0.185 | |
| ICCUBackbone=Llama-3-8B-Instruct2026.05 | — | — | — | — | — | — | 0.94 | |
| ICCUBackbone=Llama-2-7B-chat-hf2026.05 | — | — | — | — | — | — | 0.95 |