Why does Random oversampling of imbalanced dataset give worst performance with OpenPrompt?

[indunil-19]

I used a prompt classification model with a mixed template and soft verbalizer while the prompt model was frozen and only the template and verbalizer were tuned using my imbalanced dataset and the model gives 77.51% accuracy. But when I used random oversampled dataset model gives 0.52% accuracy which was really low compared to the previous result.

Can I know whether any reason behind this? (Does oversampled dataset affects prompting?)