Reward Modeling on Anthropic Helpful
72AccuracyHelpfulness RM
Evaluation results
| Method | Links | |
|---|---|---|
| Helpfulness RMFine-tuned=true2023.07 | 72 | |
| Open AssistantBackbone=DeBERTa V3 Large2023.07 | 67.7 | |
| UMM-RMBase Model=Qwen2.5-0.5B, Activated Experts=62025.11 | 67.2 | |
| SteamSHP-XLBackbone=FLAN-T5-xl2023.07 | 66.8 | |
| UMM-RMBase Model=Qwen2.5-0.5B, Activated Experts=22025.11 | 66.8 | |
| UMM-RMBase Model=Qwen2.5-0.5B, Activated Experts=42025.11 | 66 | |
| Dense RMBase Model=Qwen2.5-0.5B2025.11 | 65.8 | |
| Safety RMFine-tuned=true2023.07 | 55.4 | |
| UMM-RMBase Model=Pythia-1.4B, Activated Experts=62025.11 | 54.8 | |
| UMM-RMBase Model=Pythia-1.4B, Activated Experts=42025.11 | 53.4 | |
| UMM-RMBase Model=Pythia-1.4B, Activated Experts=22025.11 | 53 | |
| Dense RMBase Model=Pythia-1.4B2025.11 | 44.6 |