A synthetic dataset for time series super-resolution with deep learning

The increasing application of time-series analysis in fields like biomedical engineering or telecommunications emphasizes the need for high-quality data to train and evaluate advanced machine learning models. Acquiring temporal data at suitable resolutions is often limited by ethical, economic, or practical constraints. We introduce CoSiBD (Complex Signal Benchmark Dataset for Super-Resolution), a synthetic dataset designed for reproducible time-series super-resolution research. CoSiBD provides 2,500 high-resolution signals (N = 5, 000 samples each over a reference domain τ ∈ [0, 4π]) with aligned low-resolution versions at four levels (150, 250, 500, and 1,000 samples) obtained via uniform decimation. Signals are generated with diverse non-stationary behaviors through piecewise frequency modulation and spline-based amplitude envelopes, and provides both clean and noisy variants. Signals are distributed as NumPy arrays, plain text, and JSON, with comprehensive metadata describing segment structure, generation parameters, and seeds for full reproducibility. Technical validation analyzes spectral properties and reports baseline SR benchmarking and transfer experiments on EEG and speech data.