Safety-aware fine-tuning on REDORCA selected subset 2
1.37Evaluation LossOnline dynamic
Evaluation Results
| Method | Links | |
|---|---|---|
| Online dynamicsample ratio (R)=10%, G=1, backbone=LLAMA-8B2025.11 | 1.37 | |
| Onlinesample ratio (R)=10%, G=1, backbone=LLAMA-8B2025.11 | 1.45 | |
| Onlinesample ratio (R)=5%, G=1, backbone=LLAMA-8B2025.11 | 1.58 | |
| Random selectionbackbone=LLAMA-8B2025.11 | 1.63 | |
| DSIRbackbone=LLAMA-8B2025.11 | 1.67 | |
| Offline BDSbackbone=LLAMA-8B2025.11 | 1.72 | |
| LESSbackbone=LLAMA-8B2025.11 | 1.75 | |
| Direct mixingmixing ratio (ρ)=0.5, backbone=LLAMA-8B2025.11 | 1.77 | |
| Direct mixingmixing ratio (ρ)=1, backbone=LLAMA-8B2025.11 | 2.07 |