Hazard Knowledge Evaluation on WMDP
68.98AccuracyCoA
Evaluation results
| Method | Links | |
|---|---|---|
| CoABackbone=Mistral 7B2026.03 | 68.98 | |
| SFTBackbone=Llama3.1 8B2026.03 | 68.16 | |
| CoABackbone=Llama3.1 8B2026.03 | 67.76 | |
| ExtraBackbone=Llama3.1 8B2026.03 | 67.48 | |
| SFTBackbone=Mistral 7B2026.03 | 67.48 | |
| PermLMBackbone=Llama3.1 8B2026.03 | 65.17 | |
| ExtraBackbone=Mistral 7B2026.03 | 65.03 | |
| SFTBackbone=Qwen3 1.7B2026.03 | 61.77 | |
| ExtraBackbone=Qwen3 1.7B2026.03 | 61.22 | |
| PermLMBackbone=Mistral 7B2026.03 | 60.41 | |
| CoABackbone=Qwen3 1.7B2026.03 | 59.59 | |
| PermLMBackbone=Qwen3 1.7B2026.03 | 51.84 | |
| BaseBackbone=Llama3.1 8B2026.03 | 48 | |
| BaseBackbone=Mistral 7B2026.03 | 42 | |
| DenseBase Model=DeepSeek-7B, Sparsity=Dense, Evaluation Protocol=Zero-shot2025.06 | 41 | |
| DenseBase Model=LLaMA-2-7B, Sparsity=Dense, Evaluation Protocol=Zero-shot2025.06 | 39.29 | |
| sudoLMBackbone=Mistral 7B2026.03 | 35.24 | |
| sudoLMBackbone=Qwen3 1.7B2026.03 | 29.39 | |
| MaskProBase Model=DeepSeek-7B, Sparsity=2:4, Evaluation Protocol=Zero-shot2025.06 | 28.22 | |
| BaseBackbone=Qwen3 1.7B2026.03 | 28 | |
| SparsegptBase Model=DeepSeek-7B, Sparsity=2:4, Evaluation Protocol=Zero-shot2025.06 | 27.15 | |
| Pruner-ZBase Model=DeepSeek-7B, Sparsity=2:4, Evaluation Protocol=Zero-shot2025.06 | 27.07 | |
| MaskProBase Model=LLaMA-2-7B, Sparsity=2:4, Evaluation Protocol=Zero-shot2025.06 | 26.95 | |
| SparsegptBase Model=LLaMA-2-7B, Sparsity=2:4, Evaluation Protocol=Zero-shot2025.06 | 26.61 | |
| Pruner-ZBase Model=LLaMA-2-7B, Sparsity=2:4, Evaluation Protocol=Zero-shot2025.06 | 26.52 | |
| sudoLMBackbone=Llama3.1 8B2026.03 | 24.35 |