carrotcake

Automatic data quality reports, cleaning, and EDA reports for messy pandas DataFrames.

Every real dataset has the same handful of problems: inconsistent category spellings ("Low Fat" vs "low fat" vs "LOW FAT"), missing values, zeros that actually mean "missing," statistical outliers, messy column names, and duplicate rows. carrotcake finds and fixes all of it, and generates a full HTML exploratory data analysis report, in a handful of function calls.

Install

pip install carrotcake

Depends on pandas, numpy, and matplotlib (needed for eda_report's charts).

Quick start

import pandas as pd
from carrotcake import quality_report, autoclean

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

print(quality_report(df))
# carrotcake quality report — 8523 rows x 12 columns
#   [missing] Item_Weight: 5.2% missing
#   [inconsistent_categories] Item_Fat_Content: 3 variants: ['LF', 'Low Fat', 'low fat']
#   [zero_as_missing] Item_Visibility: 8.1% zero values
#   [outliers] Item_MRP: 1.8% of values are statistical outliers
#   [duplicates] <rows>: 3 duplicate rows

df_clean = autoclean(df, group_by="Item_Type")

autoclean will:

• Standardize inconsistent categorical spellings to their most frequent original form
• Treat suspiciously frequent zeros in numeric columns as missing values
• Impute missing values (group mean/mode when group_by is given, falling back to the overall column median/mode)
• Drop exact duplicate rows (run last, since standardizing values above can turn near-duplicate rows into exact duplicates)

Two more steps are available but off by default, since they're stronger, more opinionated transformations:

autoclean(df, clean_column_names=True, handle_outliers=True)

• clean_column_names — standardizes messy headers like "Item Weight " or "Sales($)" into item_weight, sales
• handle_outliers — clips statistical outliers (IQR method) to the nearest acceptable bound

Every step can be disabled individually:

autoclean(df, fix_categories=False, fix_zero_as_missing=False)

EDA reports

from carrotcake import eda_report

eda_report(df, output="report.html")

Generates a self-contained HTML report: dataset summary, missing-value table, per-column stats, a histogram for every numeric column, a correlation heatmap, and bar charts of the most common values per categorical column. Opens in any browser, no server needed.

Auditing what changed

from carrotcake import compare

print(compare(df, df_clean))
# carrotcake compare — 8523 -> 8519 rows (4 dropped)
#   [missing_fixed] Item_Weight: 443 -> 0 missing values
#   [categories_standardized] Item_Fat_Content: 5 -> 2 unique values
#   [zero_fixed] Item_Visibility: 526 -> 0 zero values

Compares aggregate column-level statistics rather than diffing individual cells, so it stays meaningful even after row dropping/reordering.

Programmatic use

report = quality_report(df)
report.to_dict()   # plain dict
report.to_json()   # JSON string, e.g. for a CI data-quality gate

Command line

carrotcake report sales.csv
carrotcake report sales.csv --json
carrotcake clean sales.csv -o sales_clean.csv --group-by Item_Type --handle-outliers
carrotcake eda sales.csv -o report.html

Why

Built after hand-writing this exact cleaning logic across multiple data analysis projects (retail sales, hotel bookings). carrotcake packages it up so it doesn't need to be rewritten for every new dataset.

License

MIT