A grammar of machine learning. Split, fit, evaluate, assess — in Python and R.

Paper · Website · Python · R

Every ML textbook distinguishes validation from test. 294 published papers don't (Kapoor & Narayanan, 2023). ml makes the mistake inexpressible.

import ml

s = ml.split(data, "churn", seed=42)

model = ml.fit(s.train, "churn", seed=42)
ml.evaluate(model, s.valid)              # iterate freely
ml.assess(model, test=s.test)            # once — second call errors

Same four verbs in R (ml_fit, ml_assess). Same Rust engine underneath. Same result.

Beyond the core four: screen algorithms, tune hyperparameters, stack ensembles, validate against deployment rules, monitor drift in production. 11 Rust-native algorithms, 38 verbs, 173 bundled datasets.

Roth, S. (2026). A Grammar of Machine Learning Workflows. doi:10.5281/zenodo.19023838

MIT. Simon Roth, 2026.