Mathematical Reasoning and Safety Preservation on GSM8K
85.22HH Safety ScoreDualSelect
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| DualSelectModel=Qwen3-4B2026.06 | 85.22 | 87.26 | 86.24 | 88.12 | |
| DualSelectModel=Llama-3-8B2026.06 | 79.48 | 84.21 | 81.85 | 80.18 | |
| TOSSModel=Qwen3-4B2026.06 | 77.04 | 75.38 | 76.21 | 87.74 | |
| DualSelectModel=Gemma-3-1B2026.06 | 76.11 | 80.28 | 78.2 | 39.34 | |
| SQSDModel=Llama-3-8B2026.06 | 75.69 | 77.64 | 76.67 | 80.06 | |
| SQSDModel=Qwen3-4B2026.06 | 75.56 | 79.05 | 77.31 | 88.03 | |
| TOSSModel=Llama-3-8B2026.06 | 73.28 | 83.81 | 78.55 | 80.24 | |
| SQSDModel=Gemma-3-1B2026.06 | 71.78 | 72.16 | 71.97 | 39.18 | |
| TOSSModel=Gemma-3-1B2026.06 | 66.37 | 75.86 | 71.12 | 38.91 | |
| LARFModel=Qwen3-4B2026.06 | 65.52 | 63.91 | 64.72 | 87.46 | |
| SEALModel=Llama-3-8B2026.06 | 64.55 | 65.89 | 65.22 | 79.68 | |
| SEALModel=Qwen3-4B2026.06 | 63.39 | 62.47 | 62.93 | 87.38 | |
| LARFModel=Llama-3-8B2026.06 | 62.13 | 67.74 | 64.94 | 79.84 | |
| LARFModel=Gemma-3-1B2026.06 | 61.24 | 59.82 | 60.53 | 38.62 | |
| SafeInstrModel=Qwen3-4B2026.06 | 60.27 | 64.61 | 62.44 | 87.21 | |
| SEALModel=Gemma-3-1B2026.06 | 58.96 | 58.41 | 58.69 | 38.45 | |
| SafeInstrModel=Llama-3-8B2026.06 | 58.32 | 60.87 | 59.6 | 79.42 | |
| RandomModel=Qwen3-4B2026.06 | 56.36 | 57.14 | 56.75 | 86.84 | |
| SafeInstrModel=Gemma-3-1B2026.06 | 55.77 | 60.06 | 57.92 | 38.1 | |
| RandomModel=Gemma-3-1B2026.06 | 54.13 | 52.91 | 53.52 | 37.52 | |
| RandomModel=Llama-3-8B2026.06 | 53.78 | 54.91 | 54.35 | 78.88 | |
| Standard SFTModel=Gemma-3-1B2026.06 | 50 | 50 | 50 | 38.37 | |
| Standard SFTModel=Qwen3-4B2026.06 | 50 | 50 | 50 | 87.95 | |
| Standard SFTModel=Llama-3-8B2026.06 | 50 | 50 | 50 | 79.73 |