ArcadeData Detective
Data Detective
Audit a dataset before the model trains on it — find every poisoned row.
Spam-filter training set
The Spam-Filter Training Set
A teammate gathered examples for a spam classifier. Some rows are poisoned: mislabeled, duplicated, or pure noise. Flag every bad row before the model trains on lies — but don't flag clean data, false alarms cost points.
Detective's notebook
- · Mislabeled: the label contradicts the obvious content.
- · Duplicate: the same example twice double-counts its pattern.
- · Outlier: noise with no meaning; it adds confusion and nothing else.
- · Clean rows stay unflagged. Real data audits punish sloppy flagging too.
Why it matters: train on poisoned rows and the model learns the lies and repeats them with confidence. Garbage in, garbage out, exactly as in the Data Is the New Recipe lesson.