List your company

Safety-aware fine-tuning on REDORCA selected subset 2

1.37Evaluation LossOnline dynamic
1.3421.5311.721.909Nov 26, 2025

Evaluation results

MethodLinks
Online dynamicsample ratio (R)=10%, G=1, backbone=LLAMA-8B2025.111.37
Onlinesample ratio (R)=10%, G=1, backbone=LLAMA-8B2025.111.45
Onlinesample ratio (R)=5%, G=1, backbone=LLAMA-8B2025.111.58
Random selectionbackbone=LLAMA-8B2025.111.63
DSIRbackbone=LLAMA-8B2025.111.67
Offline BDSbackbone=LLAMA-8B2025.111.72
LESSbackbone=LLAMA-8B2025.111.75
Direct mixingmixing ratio (ρ)=0.5, backbone=LLAMA-8B2025.111.77
Direct mixingmixing ratio (ρ)=1, backbone=LLAMA-8B2025.112.07