Synthetic Dataset Generator

Before you can train a model, you need data that actually tests it. A straight line only proves a linear model works on a straight line — it takes moons, spirals, and checkerboards to reveal where a linear boundary breaks down.

Two families of pattern

Regression patterns generate a single continuous target from one input feature:

• Linear — the textbook case, $y = ax + b$ plus noise.
• Polynomial — a quadratic bowl that a linear fit will underfit.
• Sinusoidal — a periodic signal that needs basis expansion.
• Exponential — a growth curve, often easier after a log transform.

Classification patterns generate 2D points labeled into two or more classes:

• Blobs — well-separated Gaussian clusters, linearly separable.
• Moons — two interleaving crescents; no straight line separates them.
• Circles — concentric rings, needing a radial decision boundary.
• Spiral — interleaved arms, the classic nonlinear stress test.
• Checkerboard — alternating cells with no global linear trend at all.

Try it

Switch between regression and classification, pick a pattern, and tune sample count, noise, and (for classification) the number of classes. Every dataset is generated from a seed, so hitting Regenerate gives you a fresh draw without losing reproducibility — and CSV exports exactly what's on screen.

Why this matters

Model quality claims are only as good as the data they're tested on. A classifier that looks perfect on blobs can fall apart on a spiral. Use this generator to build intuition for which geometric structure a given model family can and can't represent before you reach for real data.