Reasoning on HellaSwag (Accuracy)
76AccuracyBase
Evaluation results
| Method | Links | |
|---|---|---|
| BaseBackbone=LLaMA-2-7B, Compression Ratio=0%2025.10 | 76 | |
| LLM-PrunerBackbone=LLaMA-2-7B, Compression Ratio=20%2025.10 | 67.83 | |
| PGSVDBackbone=LLaMA-2-7B, Compression Ratio=20%2025.10 | 60.96 | |
| SliceGPTBackbone=LLaMA-2-7B, Compression Ratio=20%2025.10 | 44.28 | |
| ConSA (head-wise, single-layer)Target Sparsity (ρ)=0.50, Granularity=head-wise, Constraint Scope=single-layer, Model=1.7B2026.06 | 37.93 | |
| ConSA (layer-wise)Target Sparsity (ρ)=0.50, Granularity=layer-wise, Model=1.7B2026.06 | 36.98 | |
| Dense FATarget Sparsity (ρ)=0, Model=1.7B2026.06 | 36.35 | |
| Rule (head-wise)Target Sparsity (ρ)=0.50, Granularity=head-wise, Model=1.7B2026.06 | 34.61 | |
| Rule (layer-wise)Target Sparsity (ρ)=0.50, Granularity=layer-wise, Model=1.7B2026.06 | 34.31 | |
| ConSA (head-wise, all-layers)Target Sparsity (ρ)=0.50, Granularity=head-wise, Constraint Scope=all-layers, Model=1.7B2026.06 | 34.27 |