Free Train/Test Data Leakage Detector

TensorTurn's data leakage detector is a free tool that finds when information from your test set has bled into training, which is the single most common reason a model scores 99% in validation and fails in production. For tabular data it flags duplicate rows that appear across splits and columns suspiciously correlated with the target; for image datasets it detects the same or near-duplicate images sitting in both train and test. Upload a file, get the leaks ranked by severity, in the browser and on the free tier.

What is data leakage, and why it wrecks your metrics

Data leakage happens when your model gets to see, directly or indirectly, information it will not have at prediction time, usually because something from the test set has contaminated training. The result is validation accuracy that is far too optimistic. You ship the model, it faces genuinely unseen data, and the numbers crater. Because the model looked excellent during development, leakage is painful to diagnose after the fact, which is exactly why catching it before training is worth so much.

Types of leakage TensorTurn catches

• Train/test row leakage: identical or duplicate rows that appear in both the training and evaluation splits, so the model is effectively graded on data it memorized.
• Suspicious target correlation: columns that are almost perfectly correlated with the label. Correlation analysis surfaces these so you can check whether a feature is a legitimate signal or an accidental proxy for the answer.
• Cross-split image leakage: the same or near-duplicate image, detected with perceptual hashing and DINOv2 embeddings, present in both train and validation/test, which quietly inflates image-model accuracy.

How the detector works

• Import a CSV, Excel workbook, or image dataset (zip/rar/7z or URL).
• The detector compares splits and columns: hashing rows to find exact and duplicate overlaps, computing correlations against the target, and embedding images to catch semantic near-duplicates across splits.
• Each finding is reported with severity, the specific rows, columns, or image pairs involved, and a short explanation of the risk.
• You fix the leaks and retrain, all in one place.

The real cost of undetected leakage

Undetected leakage does not just waste a training run. It erodes trust in the whole pipeline, because every metric downstream is built on a false baseline. Teams make deployment decisions, set thresholds, and promise stakeholders numbers that were never real. A ten-second leakage scan protects every decision that depends on those metrics being honest.

A high correlation is a flag, not a verdict

TensorTurn surfaces columns that are highly correlated with the target so you can investigate them, but a strong correlation does not automatically mean leakage. Sometimes a feature is genuinely predictive. The tool gives you the signal and the context; you make the call on whether a column is a legitimate feature or a leak that has to be dropped.

Fix leakage, then train

Once you know where the leaks are, the fixes are usually straightforward: deduplicate rows, drop or re-derive leaky columns, and re-split image sets so near-duplicates stay together. You can apply reviewable transforms with the AI data editor, then describe the model you want in plain English and train it on isolated cloud GPUs, with the leakage already handled.

Scan your dataset for leakage free at https://www.tensorturn.com/login. No credit card, free tier included.