General Capability Retention on Held-out Capability Average
82.6Unweighted Average AccuracyBase
Evaluation Results
| Method | Links | |
|---|---|---|
| BaseModel backbone=Qwen3-4B-It, Fine-tuning method=Base2026.05 | 82.6 | |
| MIXSDModel backbone=Qwen3-4B-It, Fine-tuning method=MIXSD, Mixing rate λ=0.72026.05 | 79.9 | |
| MIXSDModel backbone=Qwen3-4B-It, Fine-tuning method=MIXSD, Mixing rate λ=0.52026.05 | 77.5 | |
| MIXSDModel backbone=Qwen3-4B-It, Fine-tuning method=MIXSD, Mixing rate λ=02026.05 | 76.4 | |
| MIXSDModel backbone=Qwen3-4B-It, Fine-tuning method=MIXSD, Mixing rate λ=0.32026.05 | 74.3 | |
| MIXSDModel backbone=Qwen3-8B, Fine-tuning method=MIXSD, Mixing rate λ=0.52026.05 | 72.4 | |
| BaseModel backbone=Qwen3-8B, Fine-tuning method=Base2026.05 | 72.2 | |
| MIXSDModel backbone=Qwen3-8B, Fine-tuning method=MIXSD, Mixing rate λ=0.72026.05 | 70.8 | |
| MIXSDModel backbone=Qwen3-8B, Fine-tuning method=MIXSD, Mixing rate λ=0.32026.05 | 70 | |
| MIXSDModel backbone=Qwen3-8B, Fine-tuning method=MIXSD, Mixing rate λ=02026.05 | 66.7 | |
| OPSDModel backbone=Qwen3-8B, Fine-tuning method=OPSD2026.05 | 61.8 | |
| OPSDModel backbone=Qwen3-4B-It, Fine-tuning method=OPSD2026.05 | 60.6 | |
| BaseModel backbone=Qwen3-1.7B, Fine-tuning method=Base2026.05 | 56.5 | |
| SFTModel backbone=Qwen3-8B, Fine-tuning method=SFT2026.05 | 46.1 | |
| SFTModel backbone=Qwen3-4B-It, Fine-tuning method=SFT2026.05 | 43.2 | |
| MIXSDModel backbone=Qwen3-1.7B, Fine-tuning method=MIXSD, Mixing rate λ=0.52026.05 | 40.3 | |
| MIXSDModel backbone=Qwen3-1.7B, Fine-tuning method=MIXSD, Mixing rate λ=0.32026.05 | 39.9 | |
| MIXSDModel backbone=Qwen3-1.7B, Fine-tuning method=MIXSD, Mixing rate λ=0.72026.05 | 36.1 | |
| SFTModel backbone=Qwen3-1.7B, Fine-tuning method=SFT2026.05 | 14.3 | |
| MIXSDModel backbone=Qwen3-1.7B, Fine-tuning method=MIXSD, Mixing rate λ=02026.05 | 8.6 | |
| OPSDModel backbone=Qwen3-1.7B, Fine-tuning method=OPSD2026.05 | 5.1 |