The drop from training to validation happens mainly due to overfitting and generalization gaps.

During training, the model learns patterns from data it has already seen. Validation tests whether it can apply those patterns to unseen data. If performance drops, it usually means:

• The model memorized noise instead of true patterns

• The validation data has slight distribution differences

• There is label noise or class imbalance

A small gap is normal. A large gap signals overfitting or data misalignment.