RadioMod-R16: A Simulated Small-Sample Dataset for Automatic Modulation Classification

RadioMod-R16 is a self-generated simulated benchmark dataset for small-sample automatic modulation classification (AMC). It was developed to investigate robust wireless signal representation under limited-data supervision, where only a small number of labeled samples are available for each class under each noise condition. The design of RadioMod-R16 is motivated by a practical limitation frequently encountered in non-cooperative wireless sensing and recognition: in many realistic scenarios, acquiring large-scale labeled signal data across all modulation categories and channel conditions is costly or infeasible. Existing public AMC benchmarks are valuable for large-scale supervised learning, but they are less suitable for systematically evaluating methods in explicitly constrained small-sample settings. RadioMod-R16 addresses this gap by enforcing a strict per-class, per-SNR sample budget. The dataset includes 16 mainstream digital modulation schemes: BPSK, QPSK, 8PSK, OQPSK, DQPSK, π/4-QPSK, 16QAM, 32QAM, 64QAM, 16APSK, 2FSK, 4FSK, MSK, CPFSK, GMSK, and GFSK. Each signal instance is stored as a 512-sample dual-channel I/Q baseband sequence. The dataset covers an SNR range from -10 dB to 30 dB with a step size of 2 dB. For every modulation type at every SNR point, exactly 50 samples are generated, yielding a total of 16,800 samples. To improve realism, RadioMod-R16 incorporates multiple communication-system effects during simulation, including root raised cosine pulse shaping with α = 0.35, Ricean fading, multipath ISI, random carrier frequency offset (CFO), and sampling time drift. Samples are independently generated using randomized symbol streams, channel realizations, and impairment parameters, so the dataset captures diverse waveform realizations rather than repeated deterministic patterns. RadioMod-R16 is particularly suitable for benchmarking methods that target: small-sample AMC, feature-based modulation recognition, representation learning under channel distortion, and robust classification with lightweight or non-deep architectures. Because the dataset is simulated, it should be regarded as a controlled benchmark for methodology evaluation rather than a substitute for over-the-air recordings. Its main value lies in providing a reproducible and task-oriented testbed for analyzing how representation quality, classifier design, and channel perturbations interact when labeled data are severely limited. Please cite the accompanying publication when using RadioMod-R16 in academic work.