I am trying to fine tune t5 on squad v2 dataset but even with just 100 samples, model is not improving, what i mean by that is it’s loss does decrease significantly, but Exact match or f1 score donot improve at all
Technically I thought with just 100 samples model would overfit in just no time