MLVerdict

An evidence-based decision engine for tabular machine learning.

MLVerdict does not ask you to assemble scikit-learn by hand. It takes a table, determines the learning problem, selects a metric that will not flatter a bad model, trains a constrained set of estimators, and records a verdict: which model is justified, under which assumptions, with which risks.

pip install mlverdict

PyPI · GitHub

from mlverdict import Verdict

run = Verdict(enable_hpo=True).fit("data.csv", "target")
print(run)
run.artifact().save("model.joblib")

data.csv and "target" are your path and your column name. The package does not include a sample dataset.

What we built

A supervised decision path for rectangular data (CSV / DataFrame), not a notebook template and not a neural-net trainer.

The output is a written judgment (print(run)), not a raw dict of scores. If accuracy looks strong while the rare class is never predicted, the verdict says so.

Step 1 — Classify the problem (do this before you call fit)

Machine learning on a table is not one task. First decide supervised vs unsupervised. MLVerdict automates the supervised branch. The unsupervised branch is a different class of methods; the check below tells you which one you are in.

The check

Ask one question:

After I remove a single column, can I still state the business question?

Supervised (labels exist)

You already know the correct answer on historical rows. The model’s job is to predict that same field on new rows.

How to confirm: open the file, find the column you will want in production when it is not yet known, and pass that name as target.

Unsupervised (no labels)

You are not predicting a known field. You are asking the data to describe itself.

How to confirm: there is no column you would call the answer. If you invented a target tomorrow, you would need humans or another system to label it first — until then, the problem is unsupervised.

What MLVerdict does with each

Calling fit without a target is invalid on purpose: a verdict without an outcome is not a supervised decision.

Step 2 — Install

Requires Python 3.10+.

pip install mlverdict

Optional extra candidates (same race, more boosting libraries):

pip install "mlverdict[all]"

Git clone is not part of using the library. It is only for reading source or running the project’s tests.

Step 3 — Run the supervised engine

from mlverdict import Verdict

run = Verdict(enable_hpo=True).fit("your_file.csv", "your_target_column")
print(run)

You do not select algorithms. The engine includes Logistic Regression or Ridge, Random Forest, Extra Trees, and HistGradientBoosting, then optional boosters if installed. Families that do not apply (wrong task, tiny data, missing extra) are excluded with a reason.

Step 4 — Read the verdict

print(run) is the product surface: selected model, why that metric, the leaderboard, held-out test, quality/leakage watchlist, and save instructions.

Step 5 — Persist and score new rows

from mlverdict import ModelArtifact

run.artifact().save("model.joblib")
model = ModelArtifact.load("model.joblib")
model.predict(new_rows)

The artifact is the training pipeline, not weights alone. New rows must carry the same feature columns (not the target).

Search space (transparency, not a menu)

These estimators are the candidates the engine may train. They are listed so the comparison is inspectable. You do not import them.

• Linear: Logistic Regression, Ridge
• Trees: Random Forest, Extra Trees
• Boosting: HistGradientBoosting; XGBoost, LightGBM, CatBoost when present

Development

git clone https://github.com/siva1252/ml_lib.git
cd ml_lib
pip install -e ".[dev]"
pytest

License

MIT. See LICENSE.