Training first and then doing data cleaning is very valuable

I followed the tutorial lesson 2 to try to train a bear classifier model. When I did data cleaning, I found that there was always a black bear photo in the first place.
This puzzled me, this bear was black, so I thought it was a black bear, and the classification was correct. But why was his loss value so high? Later I looked up some information and found that this was really a grizzly bear, a black grizzly bear. I found that even a simple trained model knows the difference between grizzly bear and black bears better than I do. Jeremy was right, training first and then cleaning data really brought me unexpected surprises.