Continuous targets generated from a function you know, so you can measure how much of it a model recovered. All three are non-linear on purpose: a linear model should visibly struggle where a tree ensemble does not.
Generators
Friedman: The Friedman benchmarks. The target combines products, a
squared term and linear terms, and the data carries extra features that do not affect it at all — so it tests feature selection as well as fit.
MexicanHat: The radially symmetric sinc surface: smooth, with
concentric ridges no linear model can follow.
Sine: A sine wave with configurable noise. The simplest case, and a
quick check that a regressor can fit any curve.
python
>>> from tuiml.datasets.generators.regression import Friedman
>>> data = Friedman(n_samples=500, noise=0.1, random_state=0).generate()
>>> data.X.shape[0]
500