List your company
ResearchFormerly WizWand

Safety Evaluation on XSTest (test)

95XSTest ScoreDPO + OGPSA
10.7632.6354.576.37Feb 7, 2025Apr 9, 2025Jun 9, 2025Aug 9, 2025Oct 9, 2025Dec 9, 2025Feb 8, 2026
Updated 4mo ago

Evaluation results

MethodLinks
DPO + OGPSABase Model=Llama3.1-8B-Instruct, Alignment=DPO, Strategy=OGPSA2026.0295——
DPO + OGPSABase Model=Qwen2.5-7B-Instruct, Alignment=DPO, Strategy=OGPSA2026.0294.5——
SFT-DPOBase Model=Qwen2.5-7B-Instruct, Alignment=SFT-DPO, Strategy=Baseline2026.0292.5——
SFT-DPO + OGPSABase Model=Llama3.1-8B-Instruct, Alignment=SFT-DPO, Strategy=OGPSA2026.0291.5——
SFT-DPO + LoRABase Model=Qwen2.5-7B-Instruct, Alignment=SFT-DPO, Strategy=LoRA2026.0290.5——
SFT-DPO + OGPSABase Model=Qwen2.5-7B-Instruct, Alignment=SFT-DPO, Strategy=OGPSA2026.0290.5——
SFT-DPO + General DataBase Model=Llama3.1-8B-Instruct, Alignment=SFT-DPO, Strategy=General Data2026.0289——
Instruct BaselineBase Model=Llama3.1-8B-Instruct2026.0288.5——
DPO + MergedBase Model=Llama3.1-8B-Instruct, Alignment=DPO, Strategy=Merge2026.0288——
SFT + LoRABase Model=Qwen2.5-7B-Instruct, Alignment=SFT, Strategy=LoRA2026.0288——
SFT + OGPSABase Model=Qwen2.5-7B-Instruct, Alignment=SFT, Strategy=OGPSA2026.0287.5——
SFT-DPO + MergedBase Model=Qwen2.5-7B-Instruct, Alignment=SFT-DPO, Strategy=Merge2026.0287.5——
SFT-DPO + General DataBase Model=Qwen2.5-7B-Instruct, Alignment=SFT-DPO, Strategy=General Data2026.0287.5——
SFT-DPOBase Model=Llama3.1-8B-Instruct, Alignment=SFT-DPO, Strategy=Baseline2026.0287——
SFTBase Model=Qwen2.5-7B-Instruct, Alignment=SFT, Strategy=Baseline2026.0287——
DPOBase Model=Qwen2.5-7B-Instruct, Alignment=DPO, Strategy=Baseline2026.0287——
SFT + OGPSABase Model=Llama3.1-8B-Instruct, Alignment=SFT, Strategy=OGPSA2026.0286——
SFT + LoRABase Model=Llama3.1-8B-Instruct, Alignment=SFT, Strategy=LoRA2026.0285.99——
SFT + General DataBase Model=Llama3.1-8B-Instruct, Alignment=SFT, Strategy=General Data2026.0285.5——
DPO + LoRABase Model=Qwen2.5-7B-Instruct, Alignment=DPO, Strategy=LoRA2026.0285.5——
SFTBase Model=Llama3.1-8B-Instruct, Alignment=SFT, Strategy=Baseline2026.0283——
SFT + MergedBase Model=Qwen2.5-7B-Instruct, Alignment=SFT, Strategy=Merge2026.0283——
DPO + LoRABase Model=Llama3.1-8B-Instruct, Alignment=DPO, Strategy=LoRA2026.0282.5——
SFT-DPO + MergedBase Model=Llama3.1-8B-Instruct, Alignment=SFT-DPO, Strategy=Merge2026.0282——
SFT-DPO + LoRABase Model=Llama3.1-8B-Instruct, Alignment=SFT-DPO, Strategy=LoRA2026.0281.5——
DPOBase Model=Llama3.1-8B-Instruct, Alignment=DPO, Strategy=Baseline2026.0281——
DPO + MergedBase Model=Qwen2.5-7B-Instruct, Alignment=DPO, Strategy=Merge2026.0279.5——
SFT + MergeBase Model=Llama3.1-8B-Instruct, Alignment=SFT, Strategy=Merge2026.0279——
DPO + General DataBase Model=Qwen2.5-7B-Instruct, Alignment=DPO, Strategy=General Data2026.0278.5——
SFT + General DataBase Model=Qwen2.5-7B-Instruct, Alignment=SFT, Strategy=General Data2026.0277.5——
Instruct BaselineBase Model=Qwen2.5-7B-Instruct2026.0265.5——
DPO + General DataBase Model=Llama3.1-8B-Instruct, Alignment=DPO, Strategy=General Data2026.0263——
OriginBase Model=Gemma-2-9B-IT2025.0220.4——
TELLME NT-XentBase Model=Gemma-2-9B-IT, Loss Function=NT-Xent2025.0218.4——
SFTBase Model=Gemma-2-9B-IT, Training Strategy=Supervised Fine-Tuning2025.0218——
TELLMEBase Model=Gemma-2-9B-IT2025.0214——
BaseBase Model=LLaMA3.1-8B, Adaptation Method=LoRA, Selection Budget (K)=30k2026.06—6—
BlendInPair=Q→L, Intervention Rate=35.3%2026.04——54
BlendInPair=L→G, Intervention Rate=22.9%2026.04——90
BlendInPair=Q→G, Intervention Rate=35.4%2026.04——48
DEITABase Model=LLaMA3.1-8B, Adaptation Method=LoRA, Selection Budget (K)=30k2026.06—31.5—
DOG-DBase Model=LLaMA3.1-8B, Adaptation Method=LoRA, Selection Budget (K)=30k2026.06—5.2—
KMeansBase Model=LLaMA3.1-8B, Adaptation Method=LoRA, Selection Budget (K)=30k2026.06—8.5—
LaserBase Model=LLaMA3.1-8B, Adaptation Method=LoRA, Selection Budget (K)=30k2026.06—5.2—
NUDGINGPair=Q→L, Intervention Rate=35.3%2026.04——52
NUDGINGPair=L→G, Intervention Rate=22.9%2026.04——90
NUDGINGPair=Q→G, Intervention Rate=35.4%2026.04——47
RandBase Model=LLaMA3.1-8B, Adaptation Method=LoRA, Selection Budget (K)=30k2026.06—7—
SDPOBase Model=LLaMA3.1-8B, Adaptation Method=LoRA, Selection Budget (K)=30k, Selection Protocol=Tier-3 (Full DPO training before selection)2026.06—11.7—