Unlearning Evaluation on Harry Potter Q&A
30.9Embedding Space ScoreEmbedding Space Attack
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Embedding Space AttackAttack Type=Universal, Suffix Tokens=1, Backbone=LlamaHP2024.02 | 30.9 | — | — | — | |
| Embedding Space AttackAttack Type=Universal, Suffix Tokens=5, Backbone=LlamaHP2024.02 | 30.9 | — | — | — | |
| Embedding Space AttackAttack Type=Individual, Suffix Tokens=1, Backbone=LlamaHP2024.02 | 25.5 | — | — | — | |
| Embedding Space AttackAttack Type=Universal, Suffix Tokens=20, Backbone=LlamaHP2024.02 | 25.5 | — | — | — | |
| Embedding Space AttackAttack Type=Individual, Suffix Tokens=5, Backbone=LlamaHP2024.02 | 21.8 | — | — | — | |
| Embedding Space AttackAttack Type=Individual, Suffix Tokens=20, Backbone=LlamaHP2024.02 | 20 | — | — | — | |
| Head Projection Attack (HPA)Attack Type=Individual, Backbone=LlamaHP2024.02 | — | — | 7.2 | — | |
| Probability Delta Attack (PDA)Attack Type=Individual, Backbone=LlamaHP2024.02 | — | — | — | 9 |