SynthLogic's Unweight algorithm compresses large language models by 22% while retaining 99.8% of benchmark accuracy, using a ...