Safety Alignment on Refusal Evaluation Dataset
99Refusal RateBase
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| BaseDonor=MINISTRAL-3-3B, Target=QWEN3-VL-8B, Type=REVERSE2026.01 | 99 | — | |
| BaseDonor=QWEN3-VL-2B, Target=QWEN3-VL-8B, Type=IN-FAM2026.01 | 99 | — | |
| BaseDonor=QWEN3-VL-2B, Target=MINISTRAL-3-14B, Type=CROSS-FAM2026.01 | 98 | — | |
| BaseDonor=MINISTRAL-3-3B, Target=MINISTRAL-3-14B, Type=IN-FAM2026.01 | 98 | — | |
| BaseDonor=QWEN3-VL-8B, Target=GLM-4.6V, Type=DENSE->REAS2026.01 | 97 | — | |
| BaseDonor=QWEN3-VL-8B, Target=MINISTRAL-3-3B, Type=CROSS-FAM2026.01 | 95 | — | |
| BaseDonor=QWEN3-VL-8B, Target=GPT-OSS-20B, Type=DENSE->MOE2026.01 | 94 | — | |
| BaseDonor=MINISTRAL-3-14B, Target=QWEN3-VL-2B, Type=REVERSE2026.01 | 92 | — | |
| Trajectory ReplayDonor=QWEN3-VL-8B, Target=MINISTRAL-3-3B, Type=CROSS-FAM2026.01 | 14 | 0.81 | |
| Trajectory ReplayDonor=QWEN3-VL-8B, Target=GPT-OSS-20B, Type=DENSE->MOE2026.01 | 8 | 0.88 | |
| Trajectory ReplayDonor=MINISTRAL-3-3B, Target=QWEN3-VL-8B, Type=REVERSE2026.01 | 5 | 0.92 | |
| Trajectory ReplayDonor=QWEN3-VL-2B, Target=MINISTRAL-3-14B, Type=CROSS-FAM2026.01 | 2 | 0.96 | |
| Trajectory ReplayDonor=MINISTRAL-3-14B, Target=QWEN3-VL-2B, Type=REVERSE2026.01 | 1 | 0.98 | |
| Trajectory ReplayDonor=QWEN3-VL-8B, Target=GLM-4.6V, Type=DENSE->REAS2026.01 | 1 | 0.98 | |
| Trajectory ReplayDonor=QWEN3-VL-2B, Target=QWEN3-VL-8B, Type=IN-FAM2026.01 | 0 | 0.99 | |
| Trajectory ReplayDonor=MINISTRAL-3-3B, Target=MINISTRAL-3-14B, Type=IN-FAM2026.01 | 0 | 0.99 |