Post-training Safety and Utility Alignment on Mistral-7B
0Unsafe Rate (%)CB
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| CBBackbone=Mistral-7B2026.04 | 0 | 1.04 | 100 | 2.83 | 8.16 | 18 | 70 | |
| GoalBackbone=Mistral-7B2026.04 | 0 | 9.95 | 59.7 | 8.04 | 7.33 | — | 231.5 | |
| BeamBackbone=Mistral-7B, Pruning Framework=Beam2026.04 | 1.17 | 7.13 | 46 | 7.73 | 7.13 | 18 | 69.9 | |
| GreedyBackbone=Mistral-7B, Pruning Framework=Greedy2026.04 | 1.67 | 7.12 | 46 | 7.55 | 6.96 | 18 | 70.3 | |
| One-PassBackbone=Mistral-7B, Pruning Framework=One-Pass2026.04 | 11.2 | 6.17 | 36.3 | 2.97 | 2.7 | 18 | 69 | |
| DPOBackbone=Mistral-7B2026.04 | 17 | 7.47 | 23.7 | 7.88 | 7.95 | 36 | 69.5 | |
| BaseBackbone=Mistral-7B2026.04 | 22.8 | 6.38 | 22.7 | 7.71 | 8.07 | — | 70.4 |