List your company

Alignment Reward Evaluation on OffsetBias (test)

67.94RewardSFT+DPO
19.122431.796244.4757.1438Feb 16, 2025

Evaluation results

MethodLinks
SFT+DPOModel=Qwen2.5-14B2025.0267.94
SFT+VARModel=Qwen2.5-14B2025.0266.44
VARModel=Qwen2.5-3B2025.0266.06
SFT+VARModel=Qwen2.5-1.5B2025.0265.88
DPOModel=Qwen2.5-7B2025.0265.81
SFT+VARBackbone=Llama3.1-8B2025.0265.75
SFT+VARModel=Qwen2.5-3B2025.0265.63
VARModel=Qwen2.5-14B2025.0265.63
VARModel=Qwen2.5-7B2025.0265.38
SFT+VARModel=Qwen2.5-7B2025.0265.38
VARModel=Qwen2.5-1.5B2025.0264.75
VARBackbone=Llama3.1-8B2025.0263.91
SFT+DPOBackbone=Llama3.2-3B2025.0263.81
SFT+VARBackbone=Llama3.2-3B2025.0263.72
DPOModel=Qwen2.5-14B2025.0262.94
SFT+VARBackbone=Llama2-13B2025.0262.59
SFT+VARBackbone=Llama3.2-1B2025.0261.97
VARBackbone=Llama2-13B2025.0261.44
VARBackbone=Llama3.2-3B2025.0260.88
SFT+DPOBackbone=Llama3.1-8B2025.0260.88
SFT+VARModel=Qwen2.5-0.5B2025.0260.63
VARModel=Qwen2.5-0.5B2025.0259.5
SFT+DPOBackbone=Llama3.2-1B2025.0259.09
SFT+DPOBackbone=Llama2-13B2025.0259.09
SFT+VARModel=Qwen2.5-32B-Int42025.0259.09
DPOModel=Qwen2.5-3B2025.0258.97
VARModel=Qwen2.5-32B-Int42025.0257.69
SFT+DPOModel=Qwen2.5-1.5B2025.0257.63
SFT+DPOModel=Qwen2.5-7B2025.0256.91
VARBackbone=Llama3.2-1B2025.0256.81
DPOModel=Qwen2.5-1.5B2025.0256.63
SFT+DPOModel=Qwen2.5-0.5B2025.0255.88
BaseModel=Qwen2.5-7B2025.0254.09
DPOBackbone=Llama3.2-3B2025.0253.94
DPOModel=Qwen2.5-0.5B2025.0253.44
SFT+DPOModel=Qwen2.5-3B2025.0252.63
DPOBackbone=Llama2-13B2025.0251.75
DPOBackbone=Llama3.1-8B2025.0251.38
BaseModel=Qwen2.5-14B2025.0245.16
BaseModel=Qwen2.5-3B2025.0245.06
BaseModel=Qwen2.5-0.5B2025.0240.38
BaseModel=Qwen2.5-32B-Int42025.0238.97
DPOBackbone=Llama3.2-1B2025.0237.31
DPOModel=Qwen2.5-32B-Int42025.0232.38
BaseBackbone=Llama3.1-8B2025.0230.42
SFT+DPOModel=Qwen2.5-32B-Int42025.0230.23
BaseBackbone=Llama2-13B2025.0227.05
BaseBackbone=Llama3.2-3B2025.0226.61
BaseModel=Qwen2.5-1.5B2025.0226.52
BaseBackbone=Llama3.2-1B2025.0221