ResearchBenchmarksCommonsense Reasoning on PIQA, HellaSwag, WinoGrande, ARCFollow79.4PIQA AccuracyFP1678.56878.7847979.216Mar 18, 2026Evaluation ResultsMethodMethodLinksPIQA AccuracyHellaSwag AccuracyWinoGrande AccuracyARC AccuracyAverage Commonsense Reasoning AccuracyFP16Backbone=Llama-2-7BBackbone=Llama-2-7B2026.0379.458.26453.963.9RAMPBackbone=Llama-2-7BBackbone=Llama-2-7B2026.0379.257.963.753.663.6AWQ-4Backbone=Llama-2-7BBackbone=Llama-2-7B2026.0378.857.162.352.462.7Q4_K_MBackbone=Llama-2-7BBackbone=Llama-2-7B2026.0378.656.962.152.162.4