freshdata

The explainable cleaning layer for pandas — decision-preserving data hygiene.

One call turns a messy CSV, Excel, or SQL export into analysis- and ML-ready data, and tells you exactly what it changed and why.

Documentation · Quickstart · API Reference · Changelog

What is freshdata?

freshdata is an automated data-cleaning library for Python. A rule-based decision engine profiles every column — missing ratio, dtype, skewness, cardinality, inferred role — and chooses the right action per column. Every decision carries a rationale, a risk level, and a confidence score, so nothing happens silently and nothing is left unexplained.

It fills the gap between tools that only describe data and tools that only validate it: freshdata makes the cleaning decision and shows its work, producing reproducible, auditable, ML-ready output with an audit trail you can hand to a reviewer.

Key features

• One-call cleaning — fd.clean(df) handles missing values, outliers, duplicate detection (removal is opt-in), dtype repair, and messy column names.
• Per-column decision engine — infers each column's role and applies explicit, documented rules instead of one blunt global strategy.
• Explainable by design — every action carries a rationale, risk level, and confidence score; if a NaN survives, the report says why.
• Safe defaults — never imputes an identifier, modifies a target column, or removes outliers blindly.
• pandas-first, scalable when needed — pandas + NumPy core; pass a Polars frame and get one back, with optional Polars/DuckDB/Spark execution backends for larger-than-memory data.
• CLI included — clean, plan, apply-plan, profile, learn, and trust subcommands for scripting and CI pipelines without writing Python.
• Typed and tested — fully type-hinted (py.typed), vectorized, with a 93% coverage gate enforced in CI.

Installation

pip install freshdata-cleaner

The PyPI distribution is freshdata-cleaner; the import name is freshdata.

Requires Python >= 3.9 and pandas >= 1.5. The core install depends only on pandas and NumPy; everything else is an optional extra:

pip install "freshdata-cleaner[ml,polars]"

See the installation guide for the full list of extras (domain packs, format parsers, streaming, entity resolution, and more).

Quickstart

import pandas as pd
import freshdata as fd

df = pd.read_csv("messy_export.csv")

cleaned = fd.clean(df)                               # one line
cleaned, report = fd.clean(df, return_report=True)   # ... with a full audit trail
print(report.summary())

freshdata clean report
  rows:    525 -> 500 (-25)
  columns: 7 -> 6 (-1)
  missing: 421 -> 0 cell(s)
  memory:  100.8 KB -> 89.2 KB

The same operation is available from the command line:

freshdata clean messy_export.csv -o clean.csv --report audit.json

See the quickstart guide for strategies, reports, and CLI usage.

Beyond core cleaning

Optional layers, all off by default and covered in the documentation:

• Repair plans — suggest a reviewable plan, then apply exactly the approved actions.
• Context policies — compile plain-English cleaning rules into an enforceable policy.
• Streaming — micro-batch and time-series-aware cleaning with bounded memory.
• Privacy — PII detection, masking, and jurisdiction-aware anonymization policies.
• Plugins — extend the engine with your own experts, validators, and backends.
• AI Copilot (experimental) — deterministic, offline dataset analysis that returns an explainable cleaning plan and copy-ready freshdata code; no API key required.

Documentation and examples

• Documentation site — guides, API reference, benchmarks, and honest limitations.
• examples/ — runnable, self-contained scripts, indexed in examples/README.md.

Contributing

Contributions are welcome — standard GitHub flow: fork, branch, add tests, open a pull request. CI runs ruff, mypy, and the fast pytest lane on every PR. See CONTRIBUTING.md for setup and guidelines, and CODE_OF_CONDUCT.md for community standards.

License

MIT — see LICENSE.