I am creating a Question answering system using t5-small. I trained it on a Q&A dataset with 300 rows and after training the model showed good performance on inference and outputted perfect answers. But as I finetuned the same trained model on new data the outputs were incorrect.
I tried This with Questioning the old trained model and the new finetuned model with same question but the trained model answered properly and not the finetuned one.