Reasoning on ECQA
83.96CACCMirageBD
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| MirageBDBackbone=Qwen2.5-7B, Poison Ratio=0.052026.04 | 83.96 | 99.91 | |
| BenignBackbone=Qwen2.5-7B, Poison Ratio=0.052026.04 | 83.09 | — | |
| BadChainBackbone=Qwen2.5-7B, Poison Ratio=0.052026.04 | 82.95 | 0.09 | |
| SFT+RLBackbone=Qwen2.5-7B, Poison Ratio=0.052026.04 | 82.91 | 100 | |
| DecepChainBackbone=Qwen2.5-7B, Poison Ratio=0.052026.04 | 81.27 | 99.59 | |
| BadChainBackbone=Llama3.1-8B, Poison Ratio=0.052026.04 | 79.81 | 0 | |
| MirageBDBackbone=Llama3.1-8B, Poison Ratio=0.052026.04 | 79.4 | 99.36 | |
| DecepChainBackbone=Qwen2.5-3B, Poison Ratio=0.052026.04 | 79.26 | 0 | |
| BenignBackbone=Qwen2.5-3B, Poison Ratio=0.052026.04 | 78.71 | — | |
| MirageBDBackbone=Qwen2.5-3B, Poison Ratio=0.052026.04 | 77.62 | 99.82 | |
| SFT+RLBackbone=Qwen2.5-3B, Poison Ratio=0.052026.04 | 76.62 | 97.04 | |
| BadChainBackbone=Qwen2.5-3B, Poison Ratio=0.052026.04 | 76.21 | 0 | |
| BenignBackbone=Llama3.2-3B, Poison Ratio=0.052026.04 | 74.48 | — | |
| DecepChainBackbone=Llama3.2-3B, Poison Ratio=0.052026.04 | 74.25 | 98.27 | |
| BadChainBackbone=Llama3.2-3B, Poison Ratio=0.052026.04 | 73.88 | 0 | |
| MirageBDBackbone=Llama3.2-3B, Poison Ratio=0.052026.04 | 73.43 | 99.82 | |
| BadChainBackbone=Qwen2.5-1.5B, Poison Ratio=0.052026.04 | 73.29 | 0 | |
| BenignBackbone=Llama3.1-8B, Poison Ratio=0.052026.04 | 73.2 | — | |
| MirageBDBackbone=Qwen2.5-1.5B, Poison Ratio=0.052026.04 | 72.61 | 99.45 | |
| DecepChainBackbone=Qwen2.5-1.5B, Poison Ratio=0.052026.04 | 72.42 | 99.91 | |
| BenignBackbone=Qwen2.5-1.5B, Poison Ratio=0.052026.04 | 71.51 | — | |
| SFT+RLBackbone=Llama3.2-3B, Poison Ratio=0.052026.04 | 71.1 | 93.98 | |
| SFT+RLBackbone=Qwen2.5-1.5B, Poison Ratio=0.052026.04 | 70.51 | 100 | |
| SFT+RLBackbone=Llama3.1-8B, Poison Ratio=0.052026.04 | 69.78 | 0 | |
| DecepChainBackbone=Llama3.1-8B, Poison Ratio=0.052026.04 | 69.1 | 0 |