Automated classification of cardiac abnormalities from 12-lead electrocardiograms (ECGs) using deep learning may show reduced generalization across datasets collected in different clinical settings. This study evaluated whether data augmentation improves in-domain and cross-dataset generalization of ECG classification. A 1D ResNet34-based model was trained for multi-label classification using CutOut, CutMix, MixUp, and Random. In-domain performance was evaluated by 10-fold cross-validation on the China Physiological Signal Challenge 2018 dataset (6,877 ECGs, nine classes), while cross-dataset performance was evaluated on an independent U.S. dataset containing 4,175 ECGs from seven shared classes. All augmentation methods significantly improved the in-domain F1-score, with Random augmentation achieving the highest F1-score of 0.7566, a 2.81% relative improvement over baseline. Cross-dataset performance was substantially lower and augmentation effects were method-dependent. CutOut achieved the highest external F1-score of 0.5876 (2.25% improvement), while CutOut and CutMix significantly improved over baseline; Random and MixUp did not. These findings show that in-domain augmentation gains do not necessarily translate to external datasets and emphasize the importance of cross-dataset evaluation when assessing ECG model robustness.