After "re-training" I checked again the preformance on **X_train_lt5** and **X_test_lt5** and it's accuracy decreased from ~99% to ~40%. I believe in retraining the "old examples" should be partially combined with a new ones...