A reusable Machine Learning toolkit for the full data science lifecycle. From raw data to a deployed model — loading, exploration, cleaning, encoding, modeling, interpretation, and production persistence, all in one coherent library.

---

Why nanuq?

Most data science projects re-implement the same plumbing over and over: loading files, checking for missing values, encoding categories, comparing a few models, tuning a decision threshold, then saving the result for production. That boilerplate is rarely the interesting part, and rewriting it each time invites subtle bugs and inconsistencies.

nanuq packages those recurring steps into a single, well-tested library you can install and reuse across projects. It is intentionally not a black-box AutoML tool: every function does one clear thing, and when several approaches exist for a task (imputation, scaling, handling class imbalance, encoding), nanuq offers both the atomic functions and a comparator that runs them side by side so you can make an informed choice.

Key features

• Full lifecycle coverage — 16 modules spanning loading, EDA, cleaning, encoding, scaling, transformations, feature engineering, balancing, modeling, evaluation, interpretation, plotting, reporting, and persistence.
• Comparator pattern — for tasks with multiple valid strategies, compare them on your data in a single call instead of guessing.
• Production-ready persistence — save a model and its metadata as separate, versionable artifacts, with SHA256 integrity checks and custom-threshold inference.
• No surprises — no in-place mutation of your DataFrames, no hidden global state, explicit return values.
• Documented and tested — systematic docstrings, runnable examples, 161 tests.

Installation

pip install nanuq

Or with uv:

uv pip install nanuq

Requires Python 3.11 or newer.

Quick start

import nanuq

# 1. Load and explore
df = nanuq.load_csv("data.csv")
nanuq.dataset_overview(df)

# 2. Rank features by their association with the target
ranking = nanuq.rank_features_by_target_association(df, target="churn")

# 3. Train and evaluate a model
from sklearn.ensemble import GradientBoostingClassifier
model = GradientBoostingClassifier().fit(X_train, y_train)
nanuq.evaluate_classifier(model, X_train, y_train, X_test, y_test)

# 4. Save for production (model + metadata, kept separate)
meta = nanuq.build_metadata(model, X_train.columns.tolist(), threshold=0.29)
nanuq.save_artifacts(model, meta, "artifacts/")

# 5. Reload and predict with a custom decision threshold
model, meta = nanuq.load_artifacts("artifacts/")
preds = nanuq.predict_with_threshold(
    model, X_new,
    threshold=meta["threshold"],
    feature_order=meta["feature_order"],
)

Philosophy

• Reusable. Designed to be installed as a dependency and shared across projects, not copy-pasted between notebooks.
• Comparative. When several methods exist for the same task, nanuq exposes the individual functions and a comparator that puts them side by side, so the choice is data-driven rather than habitual.
• Educational. Readable code, systematic docstrings, and worked examples — the library is meant to be learned from, not just called.
• No magic. No side effects, no in-place DataFrame mutation, predictable outputs you can reason about.

Modules

Module	Purpose
loading	Read CSV / Excel / JSON / Parquet / SQL into DataFrames
exploration	Dataset overview, column typing, quick diagnostics
cleaning	Missing-value imputation (mean, median, mode) with reports
outliers	Outlier detection (Z-score, IQR) and a comparator
encoding	One-Hot, Label, Ordinal, binary encoders and a comparator
scaling	MinMax, Standard scaling and a comparator
transformations	Log, sqrt, Box-Cox, quantile, power transforms
feature_engineering	Polynomial features, interactions, discretization
balancing	class_weight, SMOTE, undersampling for imbalanced data
bivariate	Chi², Mann-Whitney, Pearson/Spearman correlations
models	Ready-to-use linear, tree, boosted, and clustering pipelines
evaluation	Metrics, stratified CV, precision-recall curve, optimal threshold
interpretation	SHAP, permutation importance, native feature importance
persistence	Save/load models and metadata for production
plots	Boxplots, confusion matrices, importance charts
reporting	Automated EDA report generation

The comparator pattern

Rather than committing to a single model or transformation up front, nanuq lets you put options side by side and decide from the results.

# Benchmark several models on the same train/test split
results = nanuq.benchmark_models(
    {
        "logistic": LogisticRegression(),
        "random_forest": RandomForestClassifier(),
        "gradient_boosting": GradientBoostingClassifier(),
    },
    X_train, y_train, X_test, y_test,
    task="classification",
)
#   -> a DataFrame ranking each model by its metrics

# Compare a feature's distribution before and after a transformation
nanuq.compare_distribution(df["income"], nanuq.log_transform(df["income"]))

# Compare a feature's mean across target classes (quick signal check)
nanuq.compare_means_by_class(df, feature="overtime_hours", target="churn")

For atomic preprocessing you also get clear, single-purpose functions: scale_minmax, scale_standard, one_hot_encode, label_encode_column, detect_outliers_iqr, and more — no hidden defaults.

Production with persistence

The persistence module follows a deliberate convention: keep the binary model and its human-readable metadata in separate files.

from nanuq import build_metadata, save_artifacts, load_artifacts, predict_with_threshold

# Save: two separate files
meta = build_metadata(
    model, X.columns.tolist(), threshold=0.286,
    metrics={"auc": 0.83, "f1": 0.54},
)
save_artifacts(model, meta, "artifacts/")
#   artifacts/model.joblib   -> the sklearn estimator (binary)
#   artifacts/metadata.json  -> threshold, feature order, metrics (readable)

# Load + predict with the stored threshold (hash is verified automatically)
model, meta = load_artifacts("artifacts/")
preds = predict_with_threshold(
    model, X_new,
    threshold=meta["threshold"],
    feature_order=meta["feature_order"],
)

Why separate the two? Because it lets you:

• inspect metrics and configuration without loading scikit-learn (plain JSON),
• version the metadata in Git while the model stays binary,
• adjust the decision threshold without retraining or touching the model,
• detect a corrupted or swapped model file via its SHA256 hash.

Threshold optimization

On imbalanced datasets, the default 0.5 decision threshold is rarely optimal. nanuq's evaluate_classifier reports the full set of metrics (including the precision-recall trade-off) so you can pick the threshold that best fits your objective — often dramatically improving recall on the minority class.

# Full evaluation on train AND test, with the metrics you need to choose a threshold
metrics = nanuq.evaluate_classifier(model, X_train, y_train, X_test, y_test)

# Once chosen, the threshold travels with the model in metadata.json,
# and inference applies it consistently:
preds = nanuq.predict_with_threshold(
    model, X_new,
    threshold=meta["threshold"],   # e.g. 0.286 instead of 0.5
    feature_order=meta["feature_order"],
)

Storing the threshold alongside the model keeps inference consistent with the way the model was tuned.

Development

# Install in editable mode with the dev tools
uv pip install -e ".[dev]"

# Run the test suite
pytest

# Lint
ruff check src/ tests/

License

MIT — © 2026 Olivier Gruwe

---

nanuq means "polar bear" in Inuktitut. 🐻‍❄️