Data validation for ML pipelines. Break a synthetic batch and see which check notices, or check a CSV file of your own. API documentation: /docs.
A synthetic table (seven features) is split into reference, calibration and evaluation parts. One batch from the evaluation part is scored as it is and again with the fault injected. Distances use limits calibrated on clean batches for a 1% false-alarm rate per feature.
Files are processed in memory and not stored or logged. Limits: CSV, 2 MB, 50,000 rows, 200 columns. Please do not upload personal data.
Infer one from reference data, or load a schema written by mlguard infer.
Category values in an inferred schema are stored as hashes.