Dataset Drift Detector
Compare numeric and categorical distributions across two datasets.
Private by design
Your inputs stay on your device and are not sent to our servers. Optional usage analytics does not receive your files or tool inputs.
How to use Dataset Drift Detector
- Open or paste your data, or load the working sample.
- Choose columns and explicit processing rules. Review the proposed result and warnings.
- Download the chosen result dataset, JSON artifact or report; save a draft to continue later.
What are the limits of Dataset Drift Detector?
Reference dataset is the baseline. Numeric bins use baseline quantiles, duplicate boundaries collapse; population stability index uses 0.5 smoothing. Categorical total variation compares proportions with missing as its own category. Scores are descriptive, not significance tests. Numeric means exclude missing cells; nonnumeric values in selected numeric fields are rejected. Inputs stay in browser memory. No workbook is created. Table limits: 5 MB total text, 25,000 rows, 64 columns and 200,000 cells; expensive operations have stricter limits. Processing can be cancelled.