General Knowledge Preservation on HS WG IFEval MMLU Suite
17.7HS DeltaL2 Reg
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| L2 RegModel=Qwen3-4B2026.05 | 17.7 | 0.9 | 23.9 | 1.3 | 11 | |
| L2 RegModel=Llama-3-8B2026.05 | 14.6 | 1.8 | 20.6 | 3.6 | 9.3 | |
| FLOWModel=Llama-3-8B2026.05 | 13.5 | 4.7 | 22.8 | 9.1 | 10.2 | |
| WiSE-FTModel=Qwen3-4B2026.05 | 11.9 | 0.5 | 23.7 | 8.9 | 11 | |
| SFTModel=Qwen3-4B2026.05 | 11.2 | 2.5 | 17.6 | 17.3 | 10.9 | |
| SFTModel=Llama-3-8B2026.05 | 10.8 | 0.9 | 23.3 | 5.2 | 9.6 | |
| DFTModel=Llama-3-8B2026.05 | 8.6 | 0.2 | 18.9 | 11 | 9.7 | |
| TALRModel=Llama-3-8B2026.05 | 7.9 | 0.3 | 7.4 | 4.9 | 5.1 | |
| WiSE-FTModel=Llama-3-8B2026.05 | 7.7 | 0.9 | 8.1 | 1.3 | 4.5 | |
| LoRAModel=Qwen3-4B2026.05 | 7.5 | 6.7 | 2.1 | 5.6 | 2.1 | |
| FLOWModel=Qwen3-4B2026.05 | 7 | 3 | 6.4 | 10 | 6.6 | |
| LoRAModel=Llama-3-8B2026.05 | 5.8 | 6.6 | 2.9 | 6.3 | 2.1 | |
| SFT (small lr)Model=Qwen3-4B2026.05 | 5.3 | 0.4 | 4.9 | 3.1 | 0.6 | |
| STMModel=Llama-3-8B2026.05 | 5.3 | 6.9 | 4 | 0.8 | 0.8 | |
| DFTModel=Qwen3-4B2026.05 | 3.4 | 0.3 | 5.7 | 1.4 | 2 | |
| FINCHModel=Qwen3-4B2026.05 | 2.9 | 0.3 | 4.4 | 0.6 | 1.6 | |
| STMModel=Qwen3-4B2026.05 | 2.4 | 2.9 | 1.7 | 2.6 | 0.4 | |
| SFT (small lr)Model=Llama-3-8B2026.05 | 1.7 | 1.9 | 5.2 | 1.1 | 0.1 | |
| FINCHModel=Llama-3-8B2026.05 | 1.2 | 2.3 | 1.6 | 1.6 | 1.7 | |
| TALRModel=Qwen3-4B2026.05 | 0.8 | 0.3 | 2.5 | 3.8 | 0.1 | |
| BaseModel=Qwen3-4B2026.05 | 0 | 0 | 0 | 0 | 0 | |
| BaseModel=Llama-3-8B2026.05 | 0 | 0 | 0 | 0 | 0 |