Safety-aware fine-tuning on BLUEORCA (val)
0.82Evaluation LossOnline dynamic
Evaluation results
| Method | Links | |
|---|---|---|
| Online dynamicsample ratio (R)=10%, G=1, backbone=LLAMA-8B2025.11 | 0.82 | |
| Onlinesample ratio (R)=10%, G=1, backbone=LLAMA-8B2025.11 | 0.83 | |
| Onlinesample ratio (R)=5%, G=1, backbone=LLAMA-8B2025.11 | 0.84 | |
| Offline BDSbackbone=LLAMA-8B2025.11 | 0.85 | |
| Direct mixingmixing ratio (ρ)=0.5, backbone=LLAMA-8B2025.11 | 0.88 | |
| DSIRbackbone=LLAMA-8B2025.11 | 0.91 | |
| Random selectionbackbone=LLAMA-8B2025.11 | 0.92 | |
| LESSbackbone=LLAMA-8B2025.11 | 0.92 | |
| Direct mixingmixing ratio (ρ)=1, backbone=LLAMA-8B2025.11 | 0.94 |