List your company

Safety Alignment Evaluation on StrongReject SR-base

100Safety RateSFT
94.675296.057697.4498.8224May 30, 2026

Evaluation results

MethodLinks
SFTArchitecture=DeepSeek-v2-Lite2026.05100
Stair-SFTArchitecture=DeepSeek-v2-Lite2026.05100
Stair-DPOArchitecture=DeepSeek-v2-Lite2026.05100
MESAArchitecture=DeepSeek-v2-Lite2026.05100
Base(instruct)Architecture=Qwen3-30B-A3B2026.05100
SFTArchitecture=Qwen3-30B-A3B2026.05100
GRPOArchitecture=Qwen3-30B-A3B2026.05100
Stair-DPOArchitecture=Qwen3-30B-A3B2026.05100
SafeXArchitecture=Qwen3-30B-A3B2026.05100
MESAArchitecture=Qwen3-30B-A3B2026.05100
Stair-SFTArchitecture=Qwen3-30B-A3B2026.0598.72
SafeXArchitecture=DeepSeek-v2-Lite2026.0598.08
GRPOArchitecture=DeepSeek-v2-Lite2026.0595.53
Base(chat)Architecture=DeepSeek-v2-Lite2026.0594.88