Train models here, now, in this tab.
Not a video, not a mockup. Real scikit-learn, compiled to WebAssembly, training on real data with no backend. Pick a dataset and the browser runs a full cross-validated leaderboard on your own machine - the same one BreezeML would.
This is real scikit-learn compiled to WebAssembly, running in your browser with no backend and nothing to install. The leaderboard below is computed by scikit-learn directly, running the same comparison BreezeML runs for you locally. Over in the documentation, the kernel goes further and installs the real breezeml wheel from PyPI, so the snippets there call the actual library.
The kernel boots on your first pick (about 10 MB, cached after). Then eight real scikit-learn classifiers train under identical 5-fold cross-validation, right here - no server sees your click.
The kernel loads scikit-learn, pandas, and numpy - exactly three of BreezeML's four core dependencies, and all three have WebAssembly builds. That is the point of the four-dependency contract: the whole stack underneath BreezeML is small enough to run anywhere, including here. No fifth dependency, ever, or a CI test fails the build.
Read the dependency contract
pip install breezeml
from breezeml import datasets, classifiers
# the exact leaderboard you just watched
classifiers.compare(datasets.iris(), "species")Nine datasets where every trap was planted on purpose.
This tab shows you the leaderboard. The Kaggle playground shows you the failures: a churn table hiding a post-outcome leak, loans with bias baked into the labels, a campaign whose naive uplift reads $18.6 when the true effect is $12, a scoring batch that drifted after training. The generator script ships inside the dataset, so you can read exactly how each trap was planted - and twelve companion notebooks walk you through catching them, one lesson each.
Each trap links to the notebook that catches it.
churn_raw.csvtwo leaks: an ID column and the answer wearing a costumeloan_approvals.csvbias in the labels; dropping the column does not fix itab_campaign.csvconfounded targeting that flatters the campaign by 55%churn_next_quarter.csvthree columns drifted, one category never seen beforepatient_survival.csvcensored durations that break naive averagesThree more places to keep experimenting, all in the browser.
Searchable reference with runnable snippets you can edit and execute in place.
Browse the docsAn empty notebook. Bring your own data and write the pipeline yourself.
Start from nothingPoint a classifier at a dataset and read its scores, right from the docs.
Try a classifier