cleanlab documentation#

cleanlab automatically finds and fixes label issues in your ML datasets.

This reduces manual work needed to fix data errors and helps train reliable ML models on noisy real-world datasets. cleanlab has already found thousands of label errors in ImageNet, MNIST, and other popular ML benchmarking datasets, so let’s get started with yours!


1. Install cleanlab#

pip install cleanlab

2. Find label errors in your data#

cleanlab finds issues in any dataset that a classifier can be trained on. The cleanlab package works with any model by using model outputs (predicted probabilities) as input – it doesn’t depend on which model created those outputs.

If you’re using a scikit-learn-compatible model (option 1), you don’t need to train a model – you can pass the model, data, and labels into CleanLearning.find_label_issues and cleanlab will handle model training for you. If you want to use any non-sklearn-compatible model (option 2), you can input the trained model’s out-of-sample predicted probabilities into find_label_issues. Examples for both options are below.

from cleanlab.classification import CleanLearning
from cleanlab.filter import find_label_issues

# Option 1 - works with sklearn-compatible models - just input the data and labels ツ
label_issues_info = CleanLearning(clf=sklearn_compatible_model).find_label_issues(data, labels)

# Option 2 - works with ANY ML model - just input the model's predicted probabilities
ordered_label_issues = find_label_issues(
    pred_probs=pred_probs,  # out-of-sample predicted probabilities from any model

CleanLearning (option 1) also works with models from most standard ML frameworks by wrapping the model for scikit-learn compliance, e.g. huggingface/tensorflow/keras (using our KerasWrapperModel), pytorch (using skorch package), etc.

By default, find_label_issues returns a boolean mask of label issues. You can instead return the indices of potential mislabeled examples by setting return_indices_ranked_by in find_label_issues. The indices are ordered by likelihood of a label error (estimated via rank.get_label_quality_scores).


The predicted probabilities, pred_probs, from your model must be out-of-sample. Never provide predictions on the same data points used to train the model – these predictions are overfit and unsuitable for finding label errors. Details on how to compute out-of-sample predicted probabilities for your entire dataset are here.

3. Train robust models with noisy labels#

cleanlab’s CleanLearning class adapts any existing (scikit-learn compatible) classification model, clf, to a more reliable one by allowing it to train directly on partially mislabeled datasets.

When the .fit() method is called, it automatically removes any examples identified as “noisy” in the provided dataset and returns a model trained only on the clean data.

from sklearn.linear_model import LogisticRegression
from cleanlab.classification import CleanLearning

cl = CleanLearning(clf=LogisticRegression())  # any sklearn-compatible classifier, labels)

# Estimate the predictions you would have gotten if you trained without mislabeled data.
predictions = cl.predict(test_data)

4. Dataset curation: fix dataset-level issues#

cleanlab’s dataset module helps you deal with dataset-level issues by finding overlapping classes (classes to merge), rank class-level label quality (classes to keep/delete), and measure overall dataset health (to track dataset quality as you make adjustments).

The example below shows how to view all dataset-level issues in one line of code with dataset.health_summary(). Check out the dataset tutorial for more examples.

from cleanlab.dataset import health_summary

health_summary(labels, pred_probs, class_names=class_names)