Safety Evaluation on SafeBench-mini and Ch³Ef (test)
11.2SafeBench ScoreWanda
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| WandaBackbone=Qwen2.5-VL, Sparsity=50%, Evaluation Setting=zero-shot, Decoding Strategy=greedy2025.05 | 11.2 | 17.74 | 14.47 | — | |
| Wanda w/ HSRBackbone=Qwen2.5-VL, Sparsity=50%, Evaluation Setting=zero-shot, Decoding Strategy=greedy, Restoration ratio=0.020‱2025.05 | 9 | 13.03 | 11.02 | 27.4 | |
| SNIPBackbone=Qwen2.5-VL, Sparsity=50%, Evaluation Setting=zero-shot, Decoding Strategy=greedy2025.05 | 4.6 | 8.12 | 6.36 | — | |
| SNIP w/ HSRBackbone=Qwen2.5-VL, Sparsity=50%, Evaluation Setting=zero-shot, Decoding Strategy=greedy, Restoration ratio=0.150‱2025.05 | 3 | 5.34 | 4.17 | 48.88 | |
| SparseGPTBackbone=Qwen2.5-VL, Sparsity=50%, Evaluation Setting=zero-shot, Decoding Strategy=greedy2025.05 | 3 | 3.21 | 3.1 | — | |
| SparseGPT w/ HSRBackbone=Qwen2.5-VL, Sparsity=50%, Evaluation Setting=zero-shot, Decoding Strategy=greedy, Restoration ratio=0.133‱2025.05 | 2.8 | 2.56 | 2.68 | 34.43 | |
| Full ModelBackbone=Qwen2.5-VL, Sparsity=0%, Evaluation Setting=zero-shot, Decoding Strategy=greedy2025.05 | 1.4 | 2.35 | 1.88 | — |