List your company
ResearchFormerly WizWand

Reward Modeling on Anthropic Harmless

74.7AccuracySafety RM
32.5843.51554.4565.385Jul 18, 2023Dec 9, 2023May 1, 2024Sep 23, 2024Feb 14, 2025Jul 8, 2025Nov 30, 2025
Updated 8mo ago

Evaluation results

MethodLinks
Safety RMFine-tuned=true2023.0774.7
Helpfulness RMFine-tuned=true2023.0771
Open AssistantBackbone=DeBERTa V3 Large2023.0768.4
UMM-RMBase Model=Pythia-1.4B, Activated Experts=62025.1155.2
UMM-RMBase Model=Pythia-1.4B, Activated Experts=22025.1154.6
UMM-RMBase Model=Pythia-1.4B, Activated Experts=42025.1154.2
UMM-RMBase Model=Qwen2.5-0.5B, Activated Experts=62025.1150.8
UMM-RMBase Model=Qwen2.5-0.5B, Activated Experts=42025.1150.2
UMM-RMBase Model=Qwen2.5-0.5B, Activated Experts=22025.1148.8
Dense RMBase Model=Pythia-1.4B2025.1148
Dense RMBase Model=Qwen2.5-0.5B2025.1138.6
SteamSHP-XLBackbone=FLAN-T5-xl2023.0734.2