DeepOpen Intent Classification via SAM and Probability Fusion
October 23, 2026
DeepOpen improves intent classification for BANKING77 and CLINC150 datasets. A single English classifier using Sharpness-Aware Minimization (SAM) reached 94.12% accuracy on BANKING77, while probability fusion of Laya and DeBERTa-v3-large achieved 98.02% on CLINC150 in-scope classes.
HOW THIS AFFECTS YOU
●
builderYou can improve intent classification accuracy in banking domains using SAM-optimized classifiers.
●
researcherThe results suggest that SAM-based continuation offers marginal gains over historical benchmarks on BANKING77.