Accurate modulation classification is a core challenge in cognitive radio, adaptive communications, spectrum analysis, and related domains, especially under dynamic channels without transmitter knowledge. To address this need, this article presents a labeled synthetic dataset designed for wireless modulation classification under realistic propagation scenarios. The signals were generated in MATLAB by modulating randomly generated bitstreams using five digital modulation schemes: BPSK, QPSK, 16-QAM, 64-QAM, and 256-QAM. These signals were then transmitted through Rayleigh and Rician fading channels with standardized parameters, along with additional impairments to enhance realism and diversity. Each modulated signal contains 1000 symbols. A comprehensive set of features was extracted from the signals, encompassing statistical, time-domain, frequency-domain, spectrogram-based, spectral correlation-based, and image-processing-based descriptors such as BRISK, MSER, and GLCM. The dataset is organized into 10 CSV files covering two channel types (Rayleigh and Rician) across five sampling frequencies: 1 MHz, 10 MHz, 100 MHz, 500 MHz, and 1 GHz. To facilitate reproducibility and encourage further experimentation, the MATLAB scripts used for signal generation and feature extraction are also provided. This dataset serves as a valuable benchmark for developing and evaluating machine learning models in modulation classification, signal identification, and wireless communication research.