is it possible that a model performs poor on the training data but performs outstanding on the testing data?
Yes, it can happen that performance on the test data is better than performance on the training data. Essentially anything is possible, but as a general matter this situation should probably be treated as an accident or anomaly that requires further investigation. E.g. analyzing whether the statistical distributions of the two datasets are somehow different.
Have you taken any of the courses here? In the Deep Learning Specialization, Professor Ng has some detailed discussions of managing datasets in DLS Course 3 that includes issues like the one you mentioned.
Yea, I ve taken ML specialisation course where I ve completed supervised learning course and currently enrolled in unsupervised learning one
Great! I have not personally taken MLS, so Iām not sure how much Professor Ng says about this kind of issue there. But a good progression would be to take DLS once you complete MLS and, as I mentioned above, he does cover this type of issue in C2 and C3 of DLS.