Automatic Modulation Classification (AMC) plays a crucial role in modern communication systems by facilitating signal identification before demodulation. In recent years, Convolutional Neural Networks (CNNs) have been increasingly adopted for this task due to their inherent pattern recognition capabilities and robustness against noise, making them particularly suitable for the highly dynamic scenarios envisioned in future wireless communication systems. Regarding the model optimization process, the data characteristics are crucial in ensuring the model’s readiness for real-world conditions. In the current literature, most open-access datasets primarily emphasize cumulative channel and hardware impairments introduced during signal transmission and reception. However, the influence of specific signal characteristics on CNN-based modulation classifiers remains underexplored. This paper addresses this gap by presenting an experimental dataset of M-ary Quadrature Amplitude Modulation (QAM) signals with diverse features, aiming to analyze their impact on CNN training performance. The results offer valuable insights into how factors such as the number of input samples, the number of examples per modulation order, and the oversampling ratio (OSR) affect model performance and generalization.