ResearchBenchmarksTopic Classification on AG News BERT-base width (test)Follow91.52AccuracyT4-linear91.46891.481591.49591.5085Mar 5, 2026Evaluation ResultsMethodMethodLinksAccuracyT4-linearModel width (d)=768, p...Model width (d)=768, p=4, Encoder parameters=7.1 M, Compression ratio=0.25×, Peak memory=1.66 GB2026.0391.52StdModel width (d)=768, p...Model width (d)=768, p=4, Encoder parameters=28.4 M, Compression ratio=1.00×, Peak memory=1.95 GB2026.0391.47