Preference Classification on Helpfulness & Harmlessness (HH) (test)
70.4Classification AccuracyBTPO
Evaluation Results
| Method | Links | |
|---|---|---|
| BTPOBackbone=Qwen2.5-7B-Instruct2025.10 | 70.4 | |
| GRPO (pair)Backbone=Qwen2.5-7B-Instruct2025.10 | 64.7 | |
| RM-R1Backbone=Qwen2.5-7B-Instruct2025.10 | 62.7 | |
| RM-R1-finetunedBackbone=Qwen2.5-7B-Instruct, Fine-tuned=true2025.10 | 62.7 |