List your company
ResearchFormerly WizWand

Preference Alignment on Skywork

80Win Rate (W)Greedy SFT
46.7255.366472.64Sep 30, 2025
Updated 4mo ago

Evaluation results

MethodLinks
Greedy SFTBase Model=Llama 8B2025.0980—13.25
Greedy DPOBase Model=Llama 8B2025.0978—13.64
Bo2Base Model=Llama 8B2025.0976—14.15
Bo2Base Model=Qwen 4B2025.0976—9.35
Greedy SFTBase Model=Qwen 4B2025.0975—9.34
Greedy SFTBase Model=Llama 3B2025.0974—7.93
Bo2Base Model=Llama 3B2025.0974—9.04
Greedy DPOBase Model=Qwen 4B2025.0974—9.54
Bo2Base Model=Qwen 0.6B2025.0973—-5.18
Greedy SFTModel=Llama 3B, Training Protocol=full finetuning2025.09727.93—
Greedy DPOBase Model=Llama 3B2025.0972—8.45
EFTBase Model=Llama 8B2025.0972—15.57
EFTBase Model=Qwen 4B2025.0971—10.35
EFTBase Model=Llama 3B2025.0968—10.03
Greedy SFTBase Model=Qwen 0.6B2025.0966—-4.41
Greedy DPOBase Model=Qwen 0.6B2025.0966—-3.73
Bo2Model=Llama 3B, Training Protocol=full finetuning2025.096311.71—
Greedy DPOModel=Llama 3B, Training Protocol=full finetuning2025.096111.5—
Greedy SFTModel=Llama 1B, Training Protocol=full finetuning2025.0960-0.95—
EFTBase Model=Qwen 0.6B2025.0958—-1.88
Bo2Model=Llama 1B, Training Protocol=full finetuning2025.09570.47—
EFTModel=Llama 3B, Training Protocol=full finetuning2025.095612.12—
Greedy DPOModel=Llama 1B, Training Protocol=full finetuning2025.09511.12—
EFTModel=Llama 1B, Training Protocol=full finetuning2025.09482—
AADModel=Llama 1B, Training Protocol=full finetuning2025.09—1.55—
AADModel=Llama 3B, Training Protocol=full finetuning2025.09—13.4—
AADBase Model=Llama 3B2025.09——13.71
AADBase Model=Llama 8B2025.09——19.27
AADBase Model=Qwen 0.6B2025.09——-0.01
AADBase Model=Qwen 4B2025.09——14.44