SafeRLHF
Benchmarks
| Task Name | Dataset Name | SOTA Result | Trend | |
|---|---|---|---|---|
| Safety Classification | SafeRLHF | F1 Score0.94 | 48 | |
| Response Harmfulness Detection | SafeRLHF | F1 Score72.1 | 41 | |
| Harmlessness preference labeling accuracy | SafeRLHF-RMB (test) | Bench Accuracy70.6 | 15 | |
| Reward Modeling | SafeRLHF Reversed | Accuracy88.4 | 9 | |
| Reward Modeling | SafeRLHF Standard | Accuracy89.8 | 9 | |
| Safety Alignment | SafeRLHF | Win Rate83 | 8 | |
| Safety Moderation | SafeRLHF | F1 Score69.9 | 7 | |
| Constitutional AI Alignment | SafeRLHF (test) | Likert Score (5-Point)4.652 | 6 | |
| Multi-label Safety Categorization | saferlhf | Macro Accuracy48.35 | 4 | |
| Safety Alignment | SafeRLHF 30K (test) | Safety94.3 | 3 | |
| Safe and Helpful Response Generation | SafeRLHF-30K (test) | Safe Response Rate94.3 | 3 | |
| Multi-objective RLHF alignment | safeRLHF (test) | Win Rate52 | 1 |
Showing 12 of 12 rows