mlready 0.2.1

ML readiness auditor for tabular data with safe normalization and reproducible cleaning recipes

mlready

mlready is a lightweight Python library for making tabular datasets safe and ready for machine-learning pipelines.

It focuses on observation → risk detection → safe normalization, not aggressive automation.

---

Why mlready?

Real-world datasets fail ML pipelines because of:

Your current list is good but incomplete for real-world ML failures. You should add only failure modes that mlready actually addresses. No marketing fluff.

Add exactly these. Nothing more.

---

Why mlready?

Real-world datasets fail ML pipelines because of:

• numeric values stored as strings ("$1,200", "1.2M")
• inconsistent booleans (Yes / no / YES / n)
• datetime columns stored as text
• mixed types within the same column
• silent missing values ("", "NA", "null", "?")
• zero-variance and near-constant columns
• high-cardinality categorical fields
• ID-like columns that leak information
• unseen categories between train and test
• silent target leakage via derived or proxy columns
• schema drift between training and inference data

mlready helps you detect these issues early and fix only what is safe.

---

Installation

Base install (data readiness utilities):

pip install mlready

Model reporting (metrics, curves, no plots):

pip install "mlready[report]"

Model reporting with visualizations:

pip install "mlready[report,viz]"
# or
pip install "mlready[full]"

---

Core API (4 functions)

1️⃣ profile(df)

Purpose: Understand the dataset (read-only).

import mlready as mr

report = mr.profile(df)

What it provides:

• dataset shape, memory usage, duplicates

• per-column:

  • dtype and inferred logical type
  • missing counts and percentages
  • unique counts and percentages
  • top values
  • pattern hints (currency, boolean, datetime)

No data is modified.

---

2️⃣ audit(df, target=None, reference_df=None)

Purpose: Detect ML risks before training.

audit_report = mr.audit(df, target="label")

What it detects:

• ID-like columns
• zero-variance columns
• high-cardinality categoricals
• numeric / boolean / datetime stored as strings
• potential target leakage
• ghost categories (train vs test mismatch)
• schema drift

Returns structured warnings, not guesses.

---

3️⃣ apply(df, recipe=None)

Purpose: Safely normalize raw data.

clean_df, recipe, report = mr.apply(df)

What it safely applies (v0.1):

• currency / numeric string → numeric ("$1,200", "(50)", "1.2M")
• boolean strings → boolean dtype (Yes / no / YES / n)
• unambiguous datetime parsing
• safe numeric downcasting (nullable-aware)

What it does NOT do:

• no column dropping
• no encoding
• no fuzzy category merging
• no feature engineering

All actions are recorded in a reproducible recipe.

Got it. You want symmetry, not examples, and no model-specific implication.

Below is the correct, precise, non-misleading way to add report() in the same style as your other APIs.

You can paste this directly into your README.

---

4️⃣ report(model, X_train, X_test, y_train, y_test, ...)

Purpose: Evaluate classification models in a structured, pipeline-safe way.

What it provides:

• train and test classification reports

• summary metrics:

  • accuracy
  • balanced accuracy
  • precision / recall / F1 (macro, weighted)

• confusion matrix (data + optional plot)

• feature importance (when supported by the model)

• confidence analysis (probability distributions, when available)

• error analysis:

  • top misclassified rows with confidence scores

• model performance curves (optional):

  • ROC and Precision–Recall (binary)
  • calibration curve (probability-based models)
  • macro One-vs-Rest ROC-AUC (multiclass)
  • optional per-class OvR ROC curves (multiclass)

Model support scope:

• designed for sklearn-compatible classifiers
• requires predict()
• uses predict_proba() and/or decision_function() when available
• supports tree-based, linear, kernel-based, and sklearn-wrapped models

What it does NOT support:

• regression models
• clustering / anomaly detection
• multilabel (multi-hot) classification
• raw deep-learning models without sklearn-style APIs

Optional dependencies are required for reporting and visualization.

---

Design Principles

• Safety first – no silent destructive actions
• Sampling-aware – fast on large data, reliable on small data
• Pipeline-friendly – deterministic behavior via recipes
• Minimal dependencies – pure pandas / numpy

---

When to use mlready

Use mlready when:

• ingesting raw CSVs or business data
• validating train vs test consistency
• preparing data before encoding / modeling
• building reproducible ML pipelines

Not intended to replace:

• feature engineering libraries
• AutoML tools
• visualization-heavy EDA tools

---

Project Status

• Version: 0.2.1
• Stable core API
• Tests included
• Ready for production pipelines (safe mode)

---

License

MIT License

---