Split & sampling inspector

How it works

1. Choose a split method and training/validation percentages. The remaining share becomes the test partition. Use at least 10 rows.
2. Generate the fixture and inspect actual split sizes, class counts and the group overlap report.
3. Export rows with their split assignments and save the recipe. Compare methods using the same seed to keep the underlying fixture.

Different splits answer different questions

Random splitting shuffles rows. Stratification allocates each label proportionally using largest remainders. Grouped splitting allocates whole five-row groups, so row percentages can differ from the requested percentages. Chronological splitting keeps date order. All methods assign each stable row_id exactly once. The report counts groups appearing in multiple partitions and absent classes; it does not claim every downstream workflow is leakage-free.

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.

Why can a class be absent from a split?

Small partitions and rare classes cannot always satisfy every percentage and include every class. A zero-size validation partition is allowed. Fit preprocessing on training data only, and choose a splitting unit that reflects your real task. This inspector uses generated fixtures so you can see those tradeoffs without uploading private data.