felixdivo/kernel_synth_annotated

KernelSynth (annotated) One million synthetic univariate time series, each 1024 points long, drawn from a Gaussian process prior whose kernel is a random composition of up to five base kernels. This is the KernelSynth procedure from Chronos with one addition: the generating kernel is kept alongside each series. The ground-truth structure behind every series is therefore known, which makes the corpus usable for interpretability work rather than only for pretraining.… See the full description on the dataset page: https://huggingface.co/datasets/felixdivo/kernel_synth_annotated.