List your company
ResearchFormerly WizWand

LLM Pruning Evaluation on LAMBADA-Std and TriviaQA

79.3Relative Retention (Generation)SparseGPT
11.90829.40446.964.396Jun 2, 2026
Updated 5mo ago

Evaluation results

MethodLinks
SparseGPTCalibration=mix_balanced, Model Scale=LLaMA-3.1-70B, Sparsity=60% unstructured2026.0679.3
WandaCalibration=C4, Model Scale=LLaMA-3.1-70B, Sparsity=60% unstructured2026.0676.32
WandaCalibration=mix_balanced, Model Scale=LLaMA-3.1-70B, Sparsity=60% unstructured2026.0675.3
WandaCalibration=IGSP (ours), Model Scale=LLaMA-3.1-70B, Sparsity=60% unstructured2026.0670.5
SparseGPTCalibration=MetaMath, Model Scale=LLaMA-3.1-70B, Sparsity=60% unstructured2026.0669.8
SparseGPTCalibration=mix_balanced, Model Scale=LLaMA-3.1-8B, Sparsity=60% unstructured2026.0659.69
MagnitudeCalibration=None, Model Scale=LLaMA-3.1-70B, Sparsity=60% unstructured2026.0652.3
SparseGPTCalibration=MetaMath, Model Scale=LLaMA-3.1-8B, Sparsity=60% unstructured2026.0646.2
WandaCalibration=C4, Model Scale=LLaMA-3.1-8B, Sparsity=60% unstructured2026.0641.72
WandaCalibration=mix_balanced, Model Scale=LLaMA-3.1-8B, Sparsity=60% unstructured2026.0640.89
WandaCalibration=IGSP (ours), Model Scale=LLaMA-3.1-8B, Sparsity=60% unstructured2026.0634.63
MagnitudeCalibration=None, Model Scale=LLaMA-3.1-8B, Sparsity=60% unstructured2026.0614.5