Safety and Helpfulness Evaluation on RRB
1HarmRGRPO SafeSearch
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GRPO SafeSearchBackbone=Qwen-2.5-7B-Instruct2025.10 | 1 | 2.7 | |
| PPO SafeSearchBackbone=Qwen-2.5-7B-Instruct2025.10 | 2.2 | 2.69 | |
| PPO SafeSearch w/o hf.Backbone=Qwen-2.5-7B-Instruct2025.10 | 2.6 | 1.81 | |
| PPO SafeSearch w/o qr. & hf.Backbone=Qwen-2.5-7B-Instruct2025.10 | 3 | 1.84 | |
| PPO SafeSearch w/o qr.Backbone=Qwen-2.5-7B-Instruct2025.10 | 8.9 | 2.42 | |
| PPO SafeSearchBackbone=Qwen-2.5-3B-Instruct2025.10 | 14.3 | 2.57 | |
| GRPO Utility-Only Ft.Backbone=Qwen-2.5-7B-Instruct2025.10 | 23 | 2.29 | |
| BaseBackbone=Qwen-2.5-7B-Instruct2025.10 | 29.1 | 2.52 | |
| Search BlockBackbone=Qwen-2.5-7B-Instruct2025.10 | 30.5 | 2.49 | |
| Document FilterBackbone=Qwen-2.5-7B-Instruct2025.10 | 31.7 | 2.53 | |
| PPO Utility-Only Ft.Backbone=Qwen-2.5-7B-Instruct2025.10 | 32.4 | 2.55 | |
| BaseBackbone=Qwen-2.5-3B-Instruct2025.10 | 39.8 | 2.6 | |
| Document FilterBackbone=Qwen-2.5-3B-Instruct2025.10 | 49.5 | 2.6 | |
| Search BlockBackbone=Qwen-2.5-3B-Instruct2025.10 | 50.5 | 2.54 | |
| PPO Utility-Only Ft.Backbone=Qwen-2.5-3B-Instruct2025.10 | 51.4 | 2.58 |