Alignment on XSTest
0.2Refusal DeltaCE + TMKL
Evaluation Results
| Method | Links | |
|---|---|---|
| CE + TMKLBackbone=Qwen2.5-7B-Instruct, Adapter=SineLoRA2026.05 | 0.2 | |
| CE + TMKLBackbone=Qwen2.5-7B-Instruct, Adapter=LoRA2026.05 | 0.5 | |
| CE + TMKLBackbone=Qwen2.5-7B-Instruct, Adapter=DoRA2026.05 | 0.6 | |
| CE + TMKLBackbone=Qwen2.5-7B-Instruct, Adapter=RandLoRA2026.05 | 0.8 | |
| CEBackbone=Qwen2.5-7B-Instruct, Adapter=RandLoRA2026.05 | 11 | |
| CEBackbone=Qwen2.5-7B-Instruct, Adapter=SineLoRA2026.05 | 14.8 | |
| CEBackbone=Qwen2.5-7B-Instruct, Adapter=LoRA2026.05 | 15.5 | |
| CEBackbone=Qwen2.5-7B-Instruct, Adapter=DoRA2026.05 | 16 | |
| no adaptationBackbone=Qwen2.5-7B-Instruct2026.05 | 98 |