Free Mislabeled Data Finder for Image Datasets

TensorTurn's mislabeled data finder is a free tool that surfaces likely label errors in image datasets and, for each suspect, proposes a corrected label with a confidence score. It works by embedding every image with DINOv2 on a GPU and using k-nearest-neighbours: when an image's closest neighbours overwhelmingly belong to a different class than its assigned label, that image gets flagged. You upload an image dataset and get a ranked list of probable mistakes to review, on the free tier, no code.

Surface likely label errors, with a suggested fix

Every hand-labeled dataset has mistakes: a dog filed under cat, a defect marked as normal, a class boundary someone got wrong at 2am. Those wrong labels quietly cap how accurate any model can become and make your evaluation numbers untrustworthy. Finding them by re-reviewing thousands of images is impractical. The mislabeled data finder does the triage for you, pointing you at the specific images most likely to be wrong and telling you what the label probably should be.

How kNN label-error detection works

The tool embeds each image with DINOv2, a strong self-supervised vision model, so that visually and semantically similar images sit close together in embedding space. For every image it looks at the nearest neighbours. If those neighbours are dominated by a class that differs from the image's own label, that is strong evidence the label is wrong, because visually alike images should usually share a class. The finder ranks images by how confident this disagreement is.

What you get for each flagged image

• The image's current (assigned) label.
• A suggested label, taken from the majority class of its nearest neighbours.
• A confidence score for the suggestion, so you can triage the most likely errors first.
• The neighbour context that drove the flag, so the reasoning is transparent.

Why label errors matter

Label noise has an outsized effect. During training it teaches the model contradictory lessons and caps achievable accuracy; during evaluation it means correct predictions get scored as wrong, so you can't even measure progress reliably. Cleaning even a small fraction of high-confidence errors often moves metrics more than another round of hyperparameter tuning, because you are fixing the ground truth the model learns from.

Review, don't blindly trust

This is a ranked list of suspects, not an oracle. Some flagged images are genuinely mislabeled, others sit on a legitimate class boundary or are ambiguous. The confidence score and neighbour context are there so you can confirm quickly, but you make the final call and relabel or remove images yourself. That keeps a human in the loop where it matters.

Scope: image classification datasets

The mislabeled data finder is built for image datasets and auto-detects ImageFolder, YOLO, COCO, and VOC layouts. It runs as part of TensorTurn's image health check alongside duplicate, blur, exposure, and corruption detection. For tabular data, TensorTurn provides a different set of checks, focused on leakage, outliers, correlations, and mixed types, rather than kNN label-error detection.

Find the likely mislabels in your image dataset free at https://www.tensorturn.com/login. No credit card, free tier included.