Point at a clean table, get a leaderboard of models. The validation scheme, the metric, and every model's dials sit in the open, because a number you cannot defend is not a result.
Open a dataset
A clean CSV or Excel table: no gaps, features numeric. The real scikit-learn runs inside your browser (WebAssembly); your file never leaves your machine.
Continue the story
The flights data exactly as VedaForge exported it: 1,133 rows, 17 numeric columns, recipe on record. Try fare_inr as a regression target or on_time as a classification one.
Clean data only. VedaML models what is already model-ready. Gaps, text features and constants are named at the door and routed to VedaForge, the repair shop of this family; nothing is imputed or encoded behind your back.
Readiness
The roster
Every algorithm is a card: switch it on or off, then choose its mode. Auto samples seeded candidate settings from a browser-sized space, and the defaults always compete. Manual hands you the dials that matter, each with its trade-off; the fine print underneath carries everything else scikit-learn accepts, so nothing is hidden.
The leaderboard
sort metric
Every model trains inside the same harness: standardization fitted on training folds only (never on the test rows), k-fold cross-validation on the training side, and one look at the holdout at the end. The board fills in model by model and re-sorts as results land.
Model window ยท
The scored CSV is the input data with the model's outputs appended (y_pred, or y_class with per-class probabilities) plus a split column, because rows the model trained on will always flatter it: judge it on the holdout rows.
K-Means
Centroids, Euclidean distances, iterate to convergence. WCSS (within-cluster sum of squares) measures how internally coherent the clusters are, and K-Means exists to minimize it; it always falls as k grows, so the argument for a k is the ELBOW where the fall flattens. The silhouette reads separation instead of coherence, so the two can disagree, and when they do, the disagreement is the discussion.
elbow analysis up to k =
k =
Hierarchical (agglomerative)
Every row starts as its own cluster and the two closest merge, again and again; the dendrogram is the record of every merge, with height = the distance bridged. A horizontal cutoff through it IS the choice of cluster count: cut high for few coarse clusters, low for many fine ones.
linkage
clusters =
DBScan (density-based)
Two dials rule everything: ε is the radius of a point's neighbourhood, MinPts the count that makes that neighbourhood dense enough. Points become core, border or noise, and clusters grow by hopping between core neighbourhoods, which is why DBScan finds arbitrary shapes and calls outliers what they are. The k-distance curve below suggests ε: its knee is where density breaks.
MinPts
ε (eps)
Result
Run one of the three methods above; the result lands here with sizes, silhouettes, the cluster map and per-cluster profiles.
Declare the task
how the validation scheme works
The holdout method keeps a test share the training never sees; that is the honest final number. K-fold cross-validation then splits the REMAINING rows k ways, trains on k−1 parts and validates on the held part, around the circle, so the leaderboard rests on k estimates instead of one lucky split. Stratified k-fold keeps every fold's class mix equal to the whole table's, which matters exactly when classes are unequal.