Zero-shot Evaluation on MMLU, ARC-C, PIQA, WinoG, GSM8K, HellaSwag, GPQA, and RACE
60.94Average ScoreBase
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| BaseSparsity=Dense2026.02 | 60.94 | 68.25 | 55.2 | 74.27 | 67.72 | 67.5 | 73.3 | 36.6 | 44.7 | |
| Sink-AwareSparsity=0.50, Pruning Backbone=SparseGPT2026.02 | 54.58 | 64.46 | 44.97 | 70.18 | 64.09 | 53.15 | 68.35 | 32.9 | 38.58 | |
| SparseGPTSparsity=0.502026.02 | 54.4 | 64.13 | 44.88 | 70.18 | 63.54 | 52.84 | 68.72 | 32.48 | 38.42 | |
| Sink-AwareSparsity=0.50, Pruning Backbone=Wanda2026.02 | 51.68 | 60.84 | 42.83 | 68.82 | 59.75 | 49.66 | 65.15 | 30.52 | 35.85 | |
| WandaSparsity=0.502026.02 | 51.58 | 60.9 | 41.64 | 68.12 | 60.38 | 49.13 | 65.4 | 30.85 | 36.2 | |
| Sink-AwareSparsity=0.75, Pruning Backbone=SparseGPT2026.02 | 31.83 | 28.02 | 21.25 | 54.73 | 51.93 | 1.59 | 47.9 | 24.65 | 24.55 | |
| SparseGPTSparsity=0.752026.02 | 31.72 | 26.92 | 22.18 | 54.52 | 51.3 | 1.44 | 48.2 | 24.3 | 24.9 | |
| Sink-AwareSparsity=0.75, Pruning Backbone=Wanda2026.02 | 30.52 | 24.1 | 21.08 | 52.99 | 52.25 | 1.14 | 46.05 | 23.72 | 22.8 | |
| WandaSparsity=0.752026.02 | 30.38 | 23.96 | 20.65 | 53.43 | 52.17 | 1.06 | 45.12 | 23.5 | 23.15 |