Prompt recovery attack on CounterFact (test)
54Top-1 AccuracyAlphaEdit
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| AlphaEditModel=Llama-3-Instruct, N=102026.02 | 54 | 88 | 94 | 88 | |
| AlphaEditModel=GPT-J, N=1002026.02 | 52 | 90 | 100 | 89 | |
| AlphaEditModel=GPT-J, N=102026.02 | 48 | 92 | 98 | 89 | |
| AlphaEditModel=GPT-J, N=502026.02 | 47 | 88 | 100 | 89 | |
| AlphaEditModel=Llama-3-Instruct, N=1002026.02 | 45 | 77 | 87 | 87 | |
| AlphaEditModel=Llama-3-Instruct, N=502026.02 | 44 | 79 | 87 | 89 | |
| AlphaEditModel=Qwen-2.5-Instruct, N=1002026.02 | 43 | 85 | 96 | 85 | |
| AlphaEditModel=Qwen-2.5-Instruct, N=502026.02 | 36 | 80 | 96 | 85 | |
| AlphaEditModel=Qwen-2.5-Instruct, N=102026.02 | 30 | 82 | 98 | 85 |