Train-Test Split Game

Current mission

Beginner Untimed

Customer Churn

Objective:

Split quality

Not audited

Best: 0

Streak: 0

Challenge setup

Game mode

Scenario

Difficulty

Splitting strategy

Train %

Valid %

Test %

100%

Dataset size48

Class imbalance25%

Duplicate frequency10%

Groups

Random seed

Leakage difficulty

Timer

Animation

Time dependency

Protect test set after audit

Hints available

Sound effects

High contrast

Reduced motion

Import data

Interactive dataset board

Drag records between zones. Click a record to inspect it.

Ready

Shape = class · outline = group · star = suspicious record

Select a record to see its label, group, timestamp, features, and leakage clues.

Live split snapshot

0 records

Training records

Validation records

Testing records

Unassigned records

Ratio accuracy

Class balance

Group isolation

Temporal integrity

Leakage audit

Not run

Run an audit
The detector will inspect duplicates, groups, chronology, classes, preprocessing, target leakage, ratios, and test reuse.

Model performance simulation

—

Training accuracy

—

Validation accuracy

—

Testing accuracy

—

Precision

—

Recall

—

Generalisation gap

Audit the split to simulate honest performance and compare it with leakage-inflated results.

Plotly split diagnostics

Results dashboard

Measure	Result	Interpretation
Strategy	Random split	Choose a strategy matching the data structure.
Requested ratios	70/15/15	Target allocation configured by the player.
Actual ratios	0/0/0	Assign every record before final evaluation.
Leakage risk	Not audited	Run the leakage detector.
Final grade	—	No score is available yet.

Generate a challenge, allocate the records, and audit your decisions.

Progress and achievements

Mastery progress

No completed challenges yet.

Learning reference

Training, validation, and testing

Training data fits model parameters. Validation data supports tuning and model selection. Testing data provides one final independent estimate.

Stratified splitting

Preserve class proportions when labels are imbalanced. Stratification helps every split represent rare outcomes.

Group-based splitting

Keep every record from the same person, machine, household, document, or location inside one partition.

Time-series splitting

Train on earlier observations and evaluate on later observations. Random shuffling can leak future information backward.

Cross-validation

Rotate validation folds across the training data. Keep the final testing set outside the tuning loop.

Data leakage

Leakage occurs when training uses information unavailable at prediction time or learns directly from validation and testing records.

Safe preprocessing

Fit scaling, imputation, feature selection, encoding, and resampling on training data only. Apply learned transformations elsewhere.

Distribution shift

Feature or label distributions can change across time, regions, users, or devices. A realistic test set should reflect deployment.

Random seeds

A recorded seed makes random splits reproducible. Compare several seeds when the dataset is small or unstable.

Suspicious performance

Near-perfect validation results can signal duplicate records, target-derived features, future information, or repeated test-set tuning.