Open-domain Question Answering on PopQA (test)
51.26EM AccuracyOriginal (Unsafe)
Evaluation Results
| Method | Links | |
|---|---|---|
| Original (Unsafe)Backbone=Llama-3.2-3B2026.03 | 51.26 | |
| SEAL-TAGBackbone=Llama-3.2-3B2026.03 | 51.07 | |
| Few-Shot PromptingBackbone=Llama-3.2-3B2026.03 | 48.57 | |
| Eraser4RAGBackbone=Llama-3.2-3B2026.03 | 43.61 | |
| P3DeferBackbone=Llama-3.2-3B2026.03 | 31.72 | |
| EVE-Agent2026.05 | 30 | |
| PrivacyMindBackbone=Llama-3.2-3B2026.03 | 28.58 | |
| DPVoteRAGBackbone=Llama-3.2-3B2026.03 | 26.19 | |
| Dr. Zero2026.05 | 13.4 | |
| Initialsearch=disabled2026.05 | 8.1 | |
| Initialsearch=enabled2026.05 | 5.5 |