Safety-aware fine-tuning on REDORCA selected (subset 1)
1.02Evaluation LossOnline dynamic
Evaluation Results
| Method | Links | |
|---|---|---|
| Online dynamicsample ratio (R)=10%, G=1, backbone=LLAMA-8B2025.11 | 1.02 | |
| Onlinesample ratio (R)=10%, G=1, backbone=LLAMA-8B2025.11 | 1.2 | |
| Onlinesample ratio (R)=5%, G=1, backbone=LLAMA-8B2025.11 | 1.22 | |
| DSIRbackbone=LLAMA-8B2025.11 | 1.24 | |
| Random selectionbackbone=LLAMA-8B2025.11 | 1.25 | |
| Offline BDSbackbone=LLAMA-8B2025.11 | 1.25 | |
| Direct mixingmixing ratio (ρ)=1, backbone=LLAMA-8B2025.11 | 1.27 | |
| LESSbackbone=LLAMA-8B2025.11 | 1.28 | |
| Direct mixingmixing ratio (ρ)=0.5, backbone=LLAMA-8B2025.11 | 1.31 |