Shortcut Hunt
Eight models that score suspiciously well. Each one has found a shortcut in its data. Can you spot what it is really using? Guess, then watch the model retrain without its shortcut, right here in your browser. At the bottom, run the same checks on your own CSV.
The hunt
Three points for a first guess, two for a second, one for a third. Keys: 1–9 pick, R reveal, N next, H facilitator mode.
Some of the rows it learned from
What is the model really using?
| level | the shortcut | guesses | points | with shortcut | without |
|---|
Share your score
Bring your own CSV
The four hunts from the game, run automatically on your file: ID columns that track the label, columns too predictive to be true, test rows already seen in training, and a calendar doing the model's work. Nothing is uploaded.
or
Needs a column with two classes to predict (or a few classes; the biggest is then tested against the rest). Up to 3,000 rows are used.
What the checks found
How well each column predicts the target on its own (AUC)
0.5 is a coin flip, 1.0 is perfect. A single column far ahead of the pack is the pattern of a leak; a whole crowd near 1 is the pattern of an easy problem. For missing values, formats and label conflicts, see Dataset Doctor.
How it's built, results and limits: How it works →