Fine-tuning

The Laya card is direct about it: the base checkpoints sit near chance on typed decisions, 0.362 against a 0.461 majority-class floor, and the shipped 0.766 belongs to a checkpoint fine-tuned on that benchmark's own training split. The notebook runs the loop on free Kaggle T4s in a few hours.

Fine-tuning also changes what you own. Lev documents a 200,000-example recipe across 26 public datasets, and its card notes that tuning reduced summary consistency to 0.271 where the untuned backbone scored 0.826. Measure the tasks you did not train on before swapping out a hosted model.