Cleanlab

Built on the confident learning research published in JAIR in 2021, Cleanlab estimates which labels in a dataset are wrong by comparing them against a model's out-of-sample predicted probabilities, with provable guarantees on label noise estimation. The library's Datalab interface audits any dataset for label errors, outliers, near duplicates, and drift, and works with text, image, audio, and tabular data across classification, multi-label, token classification, regression, segmentation, and object detection tasks. It is deliberately model agnostic: predictions can come from PyTorch, TensorFlow, scikit-learn, XGBoost, or any framework that outputs probabilities. Installation is a plain pip, uv, or conda install on Python 3.10 or newer, everything runs locally on CPU, and the code is Apache 2.0 licensed, while the company behind it sells hosted products for automated data curation. With 11,600 GitHub stars and wide citation in data-centric AI work, it is the standard first pass for teams who suspect their training set or benchmark contains mislabeled examples.