Frequently asked questions

What is freshdata?

freshdata is an automated, explainable data-cleaning library for Python and pandas. A single call — fd.clean(df) — handles missing values, outliers, duplicates, dtype repair, and column-name normalization, and logs a rationale for every decision.

How is it different from df.fillna() / df.dropna()?

fillna/dropna apply one blunt rule to the whole frame and silently corrupt data — imputing IDs, leaking targets, or deleting meaningful outliers. freshdata infers each column's role and picks the right action per column, then explains it. If a NaN survives, the report tells you exactly why.

How is it different from sweetviz / ydata-profiling?

Profiling tools describe your data; they don't change it. freshdata both profiles (fd.profile) and cleans (fd.clean) using the same inference, so the preview matches the result.

How is it different from Great Expectations?

Great Expectations validates data against expectations but doesn't repair it. freshdata repairs data and (in the enterprise layer) can also gate on a Data Trust Score. They are complementary.

Does it work for machine-learning preprocessing?

Yes. freshdata produces leakage-aware, typed output ready for scikit-learn, XGBoost, and other ML libraries. Pass target_column= so the label is never modified, and id_columns= so identifiers are never imputed.

Will it mutate my DataFrame?

No. fd.clean returns a new DataFrame and leaves the original untouched, unless you explicitly pass preserve_original=False to reuse memory.

Does it support Polars?

Yes — install pip install "freshdata-cleaner[polars]" and pass a Polars DataFrame to fd.clean; you get a Polars DataFrame back.

How do I see what it changed?

cleaned, report = fd.clean(df, return_report=True)
print(report.summary())   # human-readable
report.to_frame()         # DataFrame
report.to_dict()          # JSON

How do I preview decisions without cleaning?

Use fd.profile(df), fd.suggest_plan(df), or fd.compare_plans(df).

Is it production-ready?

freshdata is typed (py.typed), has 800+ tests with 95%+ coverage, runs on Python 3.9–3.13, and is validated against 50 real public datasets in CI.

How do I report a bug or request a feature?

Open an issue on GitHub. For security issues, see SECURITY.md.