Sample Data Lab

How it works

1. Choose classification, regression or clustering. Set rows and features; open Settings for noise, missing values and other parameters.
2. Generate and inspect the first two features in the scatterplot. Download all rows in CSV, JSON or JSONL.
3. Export the separate answer key and report. Save the recipe for a repeatable experiment; use the split inspector to explore partitioning.

Know what the model actually generates

Classification balances original class counts within one row and shifts the first two Gaussian features by class; extra features are noise. Selected labels always change to a different class. Regression uses your coefficients and intercept, adds Gaussian target noise and optionally shifts selected targets by ±20. Clusters have circularly arranged 2D centers; injected anomalies are uniform points in a −20 to 20 square. Missing cells and perturbation rows are sampled as exact rounded counts with separate streams. Ground-truth labels and masks stay in the answer-key export, not in the feature table.

Your inputs stay with you

Your inputs and generated results stay in this browser. Settings and five recent runs use local storage; clearing browser data removes them. Export a copy to keep it. Hosting may record standard page-request logs, but generated rows and participant names are not sent to the server.

Is this data suitable for training an AI model?

It is useful for teaching, pipeline checks and controlled experiments where the generating rule matters. It is not evidence of real-world coverage or improved production performance. The report lists realized missing cells, label changes and outliers. The preview shows up to 500 points; the data export contains every row. Cross-browser floating-point transforms can differ in their final digit; retain the exported data for an exact archived fixture.