Commonsense Reasoning on HellaSwag (Accuracy and Pruning Analysis)
60AccuracyUnpruned
Evaluation results
| Method | Links | ||
|---|---|---|---|
| UnprunedBackbone=Llama-3.1-8B2026.04 | 60 | 0 | |
| SCAR-ProtBackbone=Llama-3.1-8B, Sparsity=50%, Sparsity Target=FFN channel2026.04 | 37 | 0 | |
| Act. L2Backbone=Llama-3.1-8B, Sparsity=50%, Sparsity Target=FFN channel2026.04 | 35 | 19.6 | |
| SCAR-LPBackbone=Llama-3.1-8B, Sparsity=50%, Sparsity Target=FFN channel2026.04 | 35 | 0 | |
| SCAR-ConnBackbone=Llama-3.1-8B, Sparsity=50%, Sparsity Target=FFN channel2026.04 | 34 | 0 | |
| Magnitude (channel)Backbone=Llama-3.1-8B, Sparsity=50%, Sparsity Target=FFN channel2026.04 | 27 | 41.7 | |
| Wanda (channel)Backbone=Llama-3.1-8B, Sparsity=50%, Sparsity Target=FFN channel2026.04 | 26 | 24.3 | |
| SparseGPT (channel)Backbone=Llama-3.1-8B, Sparsity=50%, Sparsity Target=FFN channel2026.04 | 22 | 34.6 |