Certified Poisoning Stability on HH-RLHF
100FTS@1DPA+ROE
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| DPA+ROEBackbone=OLMo-1B, Temperature=0.25, Shards (K)=202026.02 | 100 | 99.63 | 96.33 | 85.45 | 62.01 | 8.09 | 6.55 | 3.65 | 1.89 | 0.8 | |
| DPA+ROEBackbone=Gemma2-2B, Temperature=0.25, Shards (K)=202026.02 | 100 | 98.88 | 90.62 | 72.54 | 42.41 | 8.31 | 5.92 | 3.18 | 1.74 | 0.78 | |
| DPA+ROEBackbone=Qwen1.5-4B, Temperature=0.25, Shards (K)=202026.02 | 100 | 99.18 | 94.79 | 80.13 | 53.86 | 7.96 | 7.67 | 4.52 | 2.68 | 1.3 | |
| DPA+MSCBackbone=OLMo-1B, Temperature=0.25, Shards (K)=202026.02 | 95.14 | 84.42 | 71.57 | 58.57 | 46.28 | 7.28 | 3.89 | 1.83 | 0.77 | 0.51 | |
| DPABackbone=OLMo-1B, Temperature=0.25, Shards (K)=202026.02 | 94.64 | 83.19 | 70.21 | 58.52 | 45.54 | 6.14 | 2.82 | 1.44 | 0.74 | 0.5 | |
| DPA+MSCBackbone=Qwen1.5-4B, Temperature=0.25, Shards (K)=202026.02 | 93 | 81.38 | 70 | 51.37 | 22.12 | 7.49 | 4.35 | 2.45 | 1.3 | 0.44 | |
| DPABackbone=Qwen1.5-4B, Temperature=0.25, Shards (K)=202026.02 | 91.81 | 79.68 | 67.41 | 50.22 | 22.54 | 6.3 | 3.46 | 2.08 | 1.22 | 0.45 | |
| DPA+MSCBackbone=Gemma2-2B, Temperature=0.25, Shards (K)=202026.02 | 88.42 | 72.85 | 54.42 | 31.57 | 9.428 | 6.95 | 3.42 | 1.64 | 0.67 | 0.25 | |
| DPABackbone=Gemma2-2B, Temperature=0.25, Shards (K)=202026.02 | 86.3 | 66.22 | 45.23 | 26.78 | 9.598 | 5.75 | 2.55 | 1.26 | 0.59 | 0.25 |