Commonsense Reasoning on ARC-c, ARC-e, HellaSwag, OBQA, PIQA, Winogrande (test)
0.5341ARC-c AccuracyQwen2.5-32 B
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Qwen2.5-32 BPruning Ratio=0%, Zero-shot=true2026.01 | 0.5341 | 0.8051 | 0.6491 | 0.342 | 0.8188 | 0.753 | 0.6504 | |
| Ours(w_mix)Pruning Ratio=25%, Zero-shot=true2026.01 | 0.4667 | 0.758 | 0.57 | 0.296 | 0.7845 | 0.7285 | 0.6006 | |
| Wanda-sp(w_mix)Pruning Ratio=25%, Zero-shot=true2026.01 | 0.4224 | 0.7024 | 0.524 | 0.274 | 0.7666 | 0.6164 | 0.551 | |
| FLAP(w_mix)Pruning Ratio=25%, Zero-shot=true2026.01 | 0.4224 | 0.7285 | 0.5502 | 0.286 | 0.7802 | 0.7253 | 0.5821 | |
| Ours(w_mix)Pruning Ratio=50%, Zero-shot=true2026.01 | 0.3072 | 0.5728 | 0.3944 | 0.202 | 0.7084 | 0.614 | 0.4665 | |
| Wanda-sp(w_mix)Pruning Ratio=50%, Zero-shot=true2026.01 | 0.2423 | 0.3237 | 0.2708 | 0.156 | 0.5707 | 0.5099 | 0.3456 | |
| FLAP(w_mix)Pruning Ratio=50%, Zero-shot=true2026.01 | 0.227 | 0.3636 | 0.2943 | 0.156 | 0.6436 | 0.5107 | 0.3659 |