Cleaning & Fixing

Zedda has two complementary cleaning functions:

• zd.fix() — generate copy-paste pandas code for you to review and run yourself.
• zd.clean() — apply fixes automatically with a backup, an audit trail, and a one-call undo.

Use fix() when you want to understand and edit the fix logic. Use clean() when you trust the defaults and want a safe, reversible auto-clean.

# zd.fix() — code generation

fix() scans the dataset, detects issues, and prints pandas code grouped by issue type:

Issue category	What the generated code does
Missing Values	fillna(median) for numeric, fillna(mode()[0]) for strings
Outliers	Clip at the 99th percentile
ID Columns	Drop the column (information-free for ML)
High-Cardinality Strings	Label-encode (pd.factorize)
Constant Columns	Drop the column (zero variance)

# Generate only

import zedda as zd

zd.fix("data.csv")

Zedda prints the code to the terminal. You copy it into your notebook, edit as needed, and run it.

# Apply directly

import zedda as zd

cleaned_df = zd.fix("data.csv", apply=True)

With apply=True, Zedda executes the fixes on a pandas DataFrame in memory and returns the cleaned DataFrame. The original file on disk is not modified.

# zd.clean() — auto-clean with backup

clean() is the safe, autonomous variant. The full lifecycle:

1. Backup. Writes {path}.zedda-backup next to the original. The backup is idempotent — if a backup already exists, Zedda uses the existing one rather than overwriting it.
2. Profile. Runs a full scan to detect every auto-fixable warning.
3. Apply fixes. Applies the same logic as fix(apply=True) plus optional fuzzy typo clustering if the [clean] extra is installed.
4. Write cleaned file. Default output is the input path (overwrites the original — the backup is your safety net). Pass output= to write elsewhere.
5. Write audit trail. A JSON file at {output_dir}/{stem}_cleaning_audit.json records every fix applied, with timestamps. Audit-path traversal is checked — the audit file must live in the same directory as the output.
6. Return. Returns a pandas.DataFrame of the cleaned data (or None if the input was a file and output was a path).

# Basic usage

import zedda as zd

zd.clean("data.csv")
# data.csv is now cleaned; data.csv.zedda-backup holds the original

# Custom output path

zd.clean("data.csv", output="clean.csv")
# clean.csv is the cleaned version; data.csv.zedda-backup holds the original
# clean_cleaning_audit.json records every fix applied

# Undo

zd.clean.undo("data.csv")

Restores from the .zedda-backup. Raises ZeddaError if no backup exists.

# Cleaning a DataFrame

import pandas as pd
import zedda as zd

df = pd.read_csv("data.csv")
cleaned_df = zd.clean(df)  # returns a DataFrame; no file written

When the input is a DataFrame and output is None, no file is written — clean() returns the cleaned DataFrame directly. Pass output= to also write a file.

# The audit trail

The audit JSON records, for every fix applied:

• Column name
• Issue category (e.g. missing_values, outliers, id_column)
• Action taken (e.g. fillna_median, clip_99th_percentile, drop)
• Count of affected rows
• Timestamp

This is what makes clean() safe to use in recurring pipelines — every mutation is auditable.

# What clean() does not do

• It does not deduplicate rows. Use merge() for cross-file deduplication.
• It does not join columns from other datasets.
• It does not infer or fix datetime parsing errors (CSV datetime inference is not supported).
• It does not run the [ai] LLM. Install [clean] for fuzzy typo clustering only.

# Next steps

• Python API: fix() — full argument reference.
• Python API: clean() — full argument reference.
• Warnings — the issue categories Zedda detects.
• ML Readiness — score the cleaned dataset for ML training.