BANKING77-77: New best of 94.61% on the official test set (+0.13pp) over our previous tests 94.48%.

About this article

Hi everyone, Just wanted to share a small but hard-won milestone. After a long plateau at 94.48%, we’ve pushed the official BANKING77-77 test set (original noisy training data, strict full-train protocol) to 94.61%. Key details: +0.13pp over our previous best +0.78pp over the widely cited 93.83% baseline (Official SOTA seat at 94.94%) No test leakage — 5-fold CV on official train to freeze recipe, then retrain on 100% train data, single final test eval The model remains relatively compact (~6...