Machine Unlearning on ELUDe (retain and world)
51.3RQGuardrail
Evaluation Results
| Method | Links | |
|---|---|---|
| GuardrailBackbone=Llama-3.1-8B-Instruct2024.06 | 51.3 | |
| OriginalBackbone=Llama-3.1-8B-Instruct2024.06 | 51.2 | |
| OPT-OUTBackbone=Phi-3.5-Mini-Instruct2024.06 | 49.4 | |
| NPOBackbone=Phi-3.5-Mini-Instruct, Retention Training=true2024.06 | 49.2 | |
| DPOBackbone=Phi-3.5-Mini-Instruct, Retention Training=true2024.06 | 48.6 | |
| IDKBackbone=Phi-3.5-Mini-Instruct, Retention Training=true2024.06 | 48.4 | |
| GABackbone=Phi-3.5-Mini-Instruct, Retention Training=true2024.06 | 47.3 | |
| NPOBackbone=Llama-3.1-8B-Instruct, Retention Training=true2024.06 | 46.6 | |
| OPT-OUTBackbone=Llama-3.1-8B-Instruct2024.06 | 46.6 | |
| IDKBackbone=Llama-3.1-8B-Instruct, Retention Training=true2024.06 | 46.1 | |
| GABackbone=Llama-3.1-8B-Instruct, Retention Training=true2024.06 | 45.7 | |
| DPOBackbone=Llama-3.1-8B-Instruct, Retention Training=true2024.06 | 44.9 | |
| OriginalBackbone=Phi-3.5-Mini-Instruct2024.06 | 34.9 | |
| GuardrailBackbone=Phi-3.5-Mini-Instruct2024.06 | 26.4 | |
| IDKBackbone=Phi-3.5-Mini-Instruct, Collapsed=true2024.06 | 4.3 | |
| IDKBackbone=Llama-3.1-8B-Instruct, Collapsed=true2024.06 | 3.5 | |
| GABackbone=Llama-3.1-8B-Instruct, Collapsed=true2024.06 | 0 | |
| DPOBackbone=Llama-3.1-8B-Instruct, Collapsed=true2024.06 | 0 | |
| NPOBackbone=Llama-3.1-8B-Instruct, Collapsed=true2024.06 | 0 | |
| GABackbone=Phi-3.5-Mini-Instruct2024.06 | 0 | |
| DPOBackbone=Phi-3.5-Mini-Instruct, Collapsed=true2024.06 | 0 | |
| NPOBackbone=Phi-3.5-Mini-Instruct, Collapsed=true2024.06 | 0 |