Data diagnostics#
Use diagnose to inspect missingness, duplication and support differences
between generated data and the full data used to train the generator.
It returns a DiagnosticReport with observations you can inspect and export.
Run diagnostics#
from tabgen_eval import diagnose
diagnostics = diagnose(train=training, generated=generated, schema=schema)
print(diagnostics.summary())
Both inputs must be pandas DataFrames. You can supply a
Schema to declare column meanings.
If omitted, it is inferred from training; inspect diagnostics.schema to check
those inferred roles. Diagnostics leave the supplied rows and values unchanged.
What to look for#
Input profiles: row and column counts, missing cells and records, duplicate records, and per-column dtypes and missing counts.
Structure differences: duplicate column names, missing or extra generated columns, and different column ordering.
Categorical support: generated labels absent from training, including counts of affected cells and distinct unseen labels.
Numeric support: training and generated ranges, generated values outside the training range, nonfinite values, and cells that cannot be interpreted numerically.
Use actual training data for these support checks. A smaller real test pool can omit a category that the model saw during training; generated occurrences of that category should not be treated as unseen by the model.
Inspect and export#
summary() gives a brief overview. Use to_dataframe() to inspect individual
observations, including support differences for each column:
observations = diagnostics.to_dataframe()
print(observations)
diagnostics.to_json("data-diagnostics.json")
to_dict() provides the observations as a dictionary. to_json() returns JSON
and writes it to a file when a path is supplied.