RLHF Safety Evaluation on FFT
0.82SSSelectiveRM
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SelectiveRMBackbone=LLaMA2-7B2026.05 | 0.82 | 4.64 | |
| ILDEBackbone=LLaMA2-7B2026.05 | 0.808 | 3.05 | |
| CSGNBackbone=LLaMA2-7B2026.05 | 0.795 | 1.38 | |
| NaiveBackbone=LLaMA2-7B2026.05 | 0.784 | — | |
| SelectiveRMBackbone=Qwen2.5-7B2026.05 | 0.268 | 23.72 | |
| ILDEBackbone=Qwen2.5-7B2026.05 | 0.255 | 17.58 | |
| CSGNBackbone=Qwen2.5-7B2026.05 | 0.219 | 0.88 | |
| NaiveBackbone=Qwen2.5-7B2026.05 | 0.217 | — |