plainml

Machine learning in plain English.
From a spreadsheet to a trained, explained, deployable model in one command, or one click.

Try it in your browser · Documentation · PyPI

Point plainml at a CSV (or Excel, Parquet, JSON, a URL or a database) and name the column you want to predict. It works out whether that's classification or regression, checks the data for problems, fairly compares up to 17 models (30 with --thorough, including an optional PyTorch network), explains the winner in plain English, and saves it with an HTML report and a model card. The model is then ready to make predictions, serve as an API, ship in a Docker image or export.

It also forecasts, finds groups and unusual rows, ranks columns with 19 importance methods, and watches for drift. Prefer clicking? plainml web opens a local website: drag in a file, choose what to find out, and download whichever results you need.

Contents

Why plainml · Install · Quickstart · The website · What it can do · Models · Reports and model cards · Commands · Examples · Deploying · Trust and privacy · Python API · How it works · How plainml compares · Development

Why plainml

• No code, no setup. Blank cells, prices stored as text ("$1,200"), dates, free-text notes and ID columns are all handled automatically, and every saved model applies exactly the same preparation to new data.
• Honest scores. Models are ranked by cross-validation, the winner is checked on rows it never saw, and everything is compared against a do-nothing baseline. You get warnings for likely data leaks, imbalanced classes, overfitting, poorly calibrated probabilities, and data that simply can't predict the target.
• Explained. "Higher support_calls → 'yes' more likely (24% → 38%)", which columns matter, and why any single prediction came out the way it did.
• The whole workflow. Profile → clean → train → tune → explain → predict → watch for drift → serve → deploy, plus forecasting, clustering, anomaly detection and feature importance.
• For every level. A website, a guided mode that asks questions, one-line commands, and a Python API.
• Private by default. Everything runs on your machine with no account, and --private keeps raw data values out of reports and saved files.

Install

pip install plainml

Python 3.10 or newer. Optional extras add more:

pip install "plainml[all]"          # or pick: pip install "plainml[web,boost]"

30-second quickstart

The repository ships with example datasets in examples/:

plainml train examples/churn.csv --target churned

That prints a leaderboard and a plain-English summary, and saves a run folder such as runs/20260925-125240_churn/ with the model, report.html, model_card.md and more. Then predict on any file with the same columns:

plainml predict latest examples/churn.csv -o predictions.csv

Not sure where to start? Run plainml on its own for guided mode, or plainml web for the website.

The website

pip install "plainml[web]"
plainml web

1. Upload. Drag in a file and see every column's type and a preview of the rows.
2. Pick a task. Predict a column, forecast, find groups, find unusual rows, rank columns, check drift, profile or clean. The main options are up front; the rest are under More options.
3. Watch it run. A progress bar and a live log show each step.
4. Get the results. Headline numbers, the key findings in plain English, and the full report.
5. Download what you need. Every file the run produced is listed with its own Download button, and tables can be previewed first.

The Runs page lists everything you've run (from the website or the command line), and Predict uses any saved model on a new upload. The site works in light and dark mode and on phones. It only listens on your machine; to share it on a network, add a token: plainml web --host 0.0.0.0 --token change-me.

Try it now at plainml-tpua.vercel.app, with nothing to install.

No server needed: plainml web --export site writes the same website as static files that run plainml inside the visitor's browser. Their data never leaves their computer, and the site can be hosted for free on Vercel, GitHub Pages or a Hugging Face Static Space. See the hosting guide and the website guide.

What it can do

Models

plainml models lists every model, which tasks it handles, and whether it's installed.

† with the boost extra. ‡ with the torch extra.

• Every model is compared with a do-nothing baseline (the most common class, or the average).
• --quick runs only the fast models; --models rf,lightgbm or --exclude svm choose exactly.
• Naming a --thorough model runs it without the others (--models torch).
• Models that can't suit the data are skipped with the reason shown. Poisson regression needs a target that's never negative, for example, and slow models are skipped on large data.

The PyTorch network (pip install "plainml[torch]") is a multi-layer perceptron for tables:

• Blocks of Linear → BatchNorm → ReLU → Dropout, narrowing layer by layer.
• Trained with AdamW (weight decay) on mini-batches, with early stopping on a validation split, keeping the best weights.
• Classification uses cross-entropy (class-weighted for rare classes). Regression uses Huber loss on standardised targets.
• It behaves like any scikit-learn model, so cross-validation, ensembles, tuning (width, depth, dropout, learning rate, weight decay), saving, predict, serve and deploy all work with it.
• It trains on the CPU by default. Set PLAINML_TORCH_DEVICE=cuda or mps for a GPU.

On spreadsheet-style data, gradient boosting usually wins; the network earns its place mostly inside ensembles.

Reports and model cards

Every run writes a self-contained report.html that works offline, in light and dark mode and on a phone. For a trained model, it has:

• the model comparison
• test-set results: confusion matrix, ROC and precision-recall curves, and a calibration curve, or predicted-vs-actual for numbers
• what drives the predictions
• the data checks
• how to reproduce the run

Forecasts, clusters, anomalies, importance and drift get reports of their own.

Each trained model also gets a model_card.md: a one-page summary of what it predicts, the data it learned from, how well it does (per class, too), what drives it, its caveats, and how to use it. It's meant to travel with the model.

Commands

Every command has examples in plainml COMMAND --help, and the command reference lists every option.

Examples

# Regression, ranked by mean absolute error, fast models only
plainml train examples/house_prices.csv -t price --metric mae --quick

# Choose the models, cap the time, ignore a column, use SMOTE for rare classes
plainml train examples/churn.csv -t churned --models rf,lightgbm,xgboost --time-budget 5m --drop region --balance smote

# Several yes/no targets at once (multi-label)
plainml train data.csv -t is_spam,is_urgent

# Every model, including the extras and a stacked ensemble, with honest probabilities
plainml train examples/churn.csv -t churned --thorough --calibrate

# Squeeze more accuracy out of the last run
plainml tune latest --trials 50

# Which columns matter? Random forest importance, RFE, Boruta, SHAP... and a consensus
plainml importance examples/churn.csv -t churned --methods all

# Why did row 3 get its prediction?
plainml explain latest new_customers.csv --row 3

# Customer segments, suspicious payments, next month's sales
plainml cluster examples/customers.csv --drop member_id
plainml anomaly examples/transactions.csv --label is_fraud
plainml forecast examples/daily_sales.csv -t units_sold --horizon 30

# One forecast per store, using planned promotions and public holidays
plainml forecast examples/store_sales.csv -t sales --group store --inputs promo --country US --horizon 14

# Has this month's data drifted from what the model learned on?
plainml drift latest this_month.csv

# Predict a file too big for memory, 200,000 rows at a time
plainml predict latest huge.csv -o predictions.csv --chunk-size 200000

# Tidy up: keep only the 10 newest runs
plainml runs --prune --keep 10

A walkthrough of every example dataset is in examples/README.md.

Deploying

A REST API. plainml serve MODEL starts a FastAPI server:

Add --api-key SECRET (or set PLAINML_API_KEY) and every request except /health and /docs needs the header X-API-Key: SECRET.

A Docker image. plainml deploy MODEL writes a folder you can build and run anywhere containers run (Cloud Run, App Runner, Azure Container Apps, Fly.io, Kubernetes…):

plainml deploy latest -o deploy/churn
docker build -t churn deploy/churn
docker run -p 8000:8000 -e PLAINML_API_KEY=change-me churn

The folder holds:

• the model and its model card
• a slim Dockerfile that runs as a non-root user, with a health check
• requirements.txt, pinned to the exact library versions the model was trained with (only the ones it needs)
• a README with the commands to build, run and call it

If you installed plainml from a checkout rather than PyPI, the folder also bundles a plainml wheel, so the build doesn't need PyPI.

Files. plainml export MODEL writes ONNX (checked against the original model), for C#, Java, JavaScript or C++. --format mlflow saves an MLflow model, and plainml train --mlflow logs runs to your MLflow tracking server.

Hosting. The hosting guide covers putting the website online:

• In the browser, free: on Vercel, next to the docs (vercel.json builds both), or on a Hugging Face Static Space.
• On a server: on any container host, using the Dockerfile in hosting/docker/.

Trust and privacy

• Calibration. The report shows whether "70% sure" really means right 70% of the time (a reliability curve, the Brier score and the average gap). --calibrate fixes probabilities that are off.
• Drift. A saved model remembers what its training data looked like. plainml predict warns when new rows look clearly different, and plainml drift shows which columns moved and how much that matters.
• Private runs. With --private (on train, cluster, anomaly and importance), the report, run folder and model keep only file names, column names and summary numbers: no example rows, raw values or data paths.
• Safe loading. Saved models record their library versions, and plainml explains how to fix a mismatch. Only load model files you trust: loading a joblib/pickle file can run code.

Python API

Everything the command line does is available from Python and notebooks:

import plainml

result = plainml.train("examples/churn.csv", target="churned", quick=True)
print(result.best_model, result.cv_score)
result.leaderboard  # a DataFrame (renders as a table in Jupyter)

predictions = plainml.predict(result.run_dir, "new_customers.csv", proba=True)
why = plainml.explain(result.run_dir)
print(why.sentences)

groups = plainml.cluster("examples/customers.csv", drop=["member_id"])
forecast = plainml.forecast("examples/daily_sales.csv", "units_sold", horizon=30)
ranked = plainml.feature_importance("examples/churn.csv", "churned", methods=["rf", "rfe", "boruta"])
drift = plainml.check_drift(result.run_dir, "this_month.csv")

Saved models are ordinary scikit-learn objects that accept raw rows (strings, blanks and all):

import joblib, pandas as pd

model = joblib.load("runs/20260925-125240_churn/model.joblib")
model.predict(pd.read_csv("new_customers.csv"))

See the Python API guide.

Reproducible runs

Each run saves its settings, library versions and a fingerprint of the data. Rerun it exactly with:

plainml train --config runs/20260925-125240_churn/config.yaml

or start your own commented config with plainml init.

How it works

1. Profile. Each column is classified as number, category, date, free text, ID or constant. IDs and constants are set aside, and you get warnings about leaks, imbalance, heavy blanks and tiny datasets.

2. Split. 20% of rows are held out as a final test set that no model sees while models are chosen.

3. Compare. Every model runs inside the same pipeline:

   • fill blanks, and flag which cells were blank
   • scale numbers
   • one-hot encode categories
   • TF-IDF word features for text
   • date parts

   Each is scored with 5-fold cross-validation next to the baseline. Rare classes get weights (or SMOTE), and imbalanced yes/no problems also get a tuned decision threshold.

4. Check and explain. The winner is scored on the held-out rows, checked for calibration and overfitting, and explained with permutation importance and effect curves.

5. Save. The winner is retrained on all rows and saved with:

   • its preprocessing
   • the report and model card
   • a summary of the training data, for drift checks

More detail, including forecasting, importance and drift, is in How it works.

How plainml compares

plainml isn't the only way to get a model without writing much code. It's built for people who want to understand a result and trust it, not squeeze out the last 0.5% of accuracy. It's free, runs on your machine, needs no account, and says what it found in plain English.

Single-purpose tools also do parts of this well: ydata-profiling for profiling, Evidently for drift, Prophet and StatsForecast for forecasting, and mlxtend and BorutaPy for feature selection.

When to reach for something else: AutoGluon when accuracy is all that matters; H2O for data that doesn't fit on one machine; PyCaret if you live in notebooks and want to customise every step; an enterprise platform when a team needs approvals, monitoring and access control.

Development

git clone https://github.com/pranay-obla/plainml && cd plainml
pip install -e ".[all,dev]"
pre-commit install
pytest

CI runs the tests on every Python from 3.10 to 3.14 on Linux, macOS and Windows. It also tests with every optional extra (PyTorch included), against the oldest supported library versions, and weekly against the newest releases, so a library update can't silently break plainml.

Publishing a GitHub release sends the package to PyPI (see RELEASING.md), and the documentation in docs/ can be hosted on Vercel or GitHub Pages.

Author

@pranay-obla

plainml grew out of curdrice, a hackathon project by @hrishitb, @pranayobla, @shriharik and @ankitthomas.

Acknowledgements

Built on scikit-learn, pandas, NumPy, Rich, Click, questionary and joblib, with optional XGBoost, LightGBM, CatBoost, PyTorch, Optuna, SHAP, imbalanced-learn, FastAPI, skl2onnx, holidays and MLflow.

MIT licensed.