Answers to frequently asked questions about the cleanlab open-source package.

The code snippets in this FAQ come from a fully executable notebook you can run via Colab or locally by downloading it here.

What data can cleanlab detect issues in?#

Currently, cleanlab can be used to detect label issues in any classification dataset, including those involving: multiple annotators per example (multi-annotator), or multiple labels per example (multi-label). This includes data from any modality such as: image, text, tabular, audio, etc. For text data, cleanlab also supports NLP tasks like entity recognition in which each word is individually labeled (token classification). We’re working to add support for all other common supervised learning tasks. If you have a particular task in mind, let us know!

How do I format classification labels for cleanlab?#

cleanlab only works with integer-encoded labels in the range {0,1, ... K-1} where K = number_of_classes. The labels array should only contain integer values in the range {0, K-1} and be of shape (N,) where N = total_number_of_data_points. Do not pass in labels where some classes are entirely missing or are extremely rare, as cleanlab may not perform as expected. It is better to remove such classes entirely from the dataset first (also dropping the corresponding dimensions from pred_probs and then renormalizing it).

Text or string labels should to be mapped to integers for each possible value. For example if your original data labels look like this: ["dog", "dog", "cat", "mouse", "cat"], you should feed them to cleanlab like this: labels = [1,1,0,2,0] and keep track of which integer uniquely represents which class (classes were ordered alphabetically in this example).

One-hot encoded labels should be integer-encoded by finding the argmax along the one-hot encoded axis. An example of what this might look like is shown below.

import numpy as np

# This example arr has 4 labels (one per data point) where
# each label can be one of 3 possible classes

arr  = np.array([[0,1,0],[1,0,0],[0,0,1],[1,0,0]])
labels_proper_format = np.argmax(arr, axis=1)  # How labels should be formatted when passed into the model

How do I infer the correct labels for examples cleanlab has flagged?#

If you have a classifier that is compatible with CleanLearning (i.e. follows the sklearn API), here’s an easy way to see predicted labels alongside the label issues:

cl = cleanlab.classification.CleanLearning(your_classifier)
issues_dataframe = cl.find_label_issues(data, labels)

Alternatively if you have already computed out-of-sample predicted probabilities (pred_probs) from a classifier:

cl = cleanlab.classification.CleanLearning()
issues_dataframe = cl.find_label_issues(X=None, labels=labels, pred_probs=pred_probs)

Otherwise if you have already found issues via:

issues = cleanlab.filter.find_label_issues(labels, pred_probs)

then you can see your trained classifier’s class prediction for each flagged example like this:

class_predicted_for_flagged_examples = pred_probs[issues].argmax(axis=1)

Here you can see the classifier’s class prediction for every example via:

class_predicted_for_all_examples = pred_probs.argmax(axis=1)

We caution against just blindly taking the predicted label for granted, many of these suggestions may be wrong! You will be able to produce a much better version of your dataset interactively using Cleanlab Studio, which helps you efficiently fix issues like this in large datasets.

How should I handle label errors in train vs. test data?#

If you do not address label errors in your test data, you may not even know when you have produced a better ML model because the evaluation is too noisy. For the best-trained models and most reliable evaluation of them, you should fix label errors in both training and testing data.

To do this efficiently, first use cleanlab to automatically find label issues in both sets. You can simply merge these two sets into one larger dataset and run cross-validation training + find_label_issues() on the merged datataset. Calling the CleanLearning.find_label_issues() method on your merged dataset does both these steps for you with any scikit-learn compatible classifier you choose.

After finding label issues, be wary about auto-correcting the labels for test examples (as cautioned against above). Instead manually fix the labels for your test data via careful review of the flagged issues. You can use Cleanlab Studio to fix labels efficiently.

Auto-correcting labels for your training data is fair game, which should improve ML performance (if properly evaluated with clean test labels). You can boost ML performance further by manually fixing the training examples flagged with label issues, as demonstrated in this article:

Handling Mislabeled Tabular Data to Improve Your XGBoost Model

How can I find label issues in big datasets with limited memory?#

For a dataset with many rows and/or classes, there are more efficient methods in the label_issues_batched module. These methods read data in mini-batches and you can reduce the batch_size to control how much memory they require. Below is an example of how to use the find_label_issues_batched() method from this module, which can load mini-batches of data from labels, pred_probs saved as .npy files on disk. You can also run this method on Zarr arrays loaded from .zarr files. Try playing with the n_jobs argument for further multiprocessing speedups. If you need greater flexibility, check out the LabelInspector class from this module.

# We'll assume your big arrays of labels, pred_probs have been saved to file like this:
from tempfile import mkdtemp
import os.path as path

labels_file = path.join(mkdtemp(), "labels.npy")
pred_probs_file = path.join(mkdtemp(), "pred_probs.npy")
np.save(labels_file, labels)
np.save(pred_probs_file, pred_probs)

# Code to find label issues by loading data from file in batches:
from cleanlab.experimental.label_issues_batched import find_label_issues_batched

batch_size = 10000  # for efficiency, set this to as large of a value as your memory can handle

# Indices of examples with label issues, sorted by label quality score (most severe to least severe):
indices_of_examples_with_issues = find_label_issues_batched(
    labels_file=labels_file, pred_probs_file=pred_probs_file, batch_size=batch_size
mmap-loaded numpy arrays have: 50 examples, 3 classes
Total number of examples whose labels have been evaluated: 50

To use less memory and get results faster if your dataset has many classes: Try merging the rare classes into a single “Other” class before you find label issues. The resulting issues won’t be affected much since cleanlab anyway does not have enough data to accurately diagnose label errors in classes that are rarely seen. To do this, you should aggregate all the probability assigned to the rare classes in pred_probs into a single new dimension of pred_probs_merged (where this new array no longer has columns for the rare classes). Here is a function that does this for you, which you can also modify as needed:

from cleanlab.internal.util import value_counts  # use this to count how often each class occurs in labels

def merge_rare_classes(labels, pred_probs, count_threshold = 10):
    Returns: labels, pred_probs after we merge all rare classes into a single 'Other' class.
    Merged pred_probs has less columns. Rare classes are any occuring less than `count_threshold` times.
    Also returns: `class_mapping_orig2new`, a dict to map new classes in merged labels back to classes
    in original labels, useful for interpreting outputs from `dataset.heath_summary()` or `count.confident_joint()`.
    num_classes = pred_probs.shape[1]
    num_examples_per_class = value_counts(labels, num_classes=num_classes)
    rare_classes = [c for c in range(num_classes) if num_examples_per_class[c] < count_threshold]
    if len(rare_classes) < 1:
        raise ValueError("No rare classes found at the given `count_threshold`, merging is unnecessary unless you increase it.")

    num_classes_merged = num_classes - len(rare_classes) + 1  # one extra class for all the merged ones
    other_class = num_classes_merged - 1
    labels_merged = labels.copy()
    class_mapping_orig2new = {}  # key = original class in `labels`, value = new class in `labels_merged`
    new_c = 0
    for c in range(num_classes):
        if c in rare_classes:
            class_mapping_orig2new[c] = other_class
            class_mapping_orig2new[c] = new_c
            new_c += 1
        labels_merged[labels == c] = class_mapping_orig2new[c]

    merged_prob = np.sum(pred_probs[:, rare_classes], axis=1, keepdims=True)  # total probability over all merged classes for each example
    pred_probs_merged = np.hstack((np.delete(pred_probs, rare_classes, axis=1), merged_prob))  # assumes new_class is as close to original_class in sorted order as is possible after removing the merged original classes
    # check a few rows of probabilities after merging to verify they still sum to 1:
    num_check = 1000  # only check a few rows for efficiency
    ones_array_ref = np.ones(min(num_check,len(pred_probs)))
    if np.isclose(np.sum(pred_probs[:num_check], axis=1), ones_array_ref).all() and (not np.isclose(np.sum(pred_probs_merged[:num_check], axis=1), ones_array_ref).all()):
        raise ValueError("merged pred_probs do not sum to 1 in each row, check that merging was correctly done.")

    return (labels_merged, pred_probs_merged, class_mapping_orig2new)
from cleanlab.filter import find_label_issues  # can alternatively use find_label_issues_batched() shown above

labels_merged, pred_probs_merged, class_mapping_orig2new = merge_rare_classes(labels, pred_probs, count_threshold=5)
examples_w_issues = find_label_issues(labels_merged, pred_probs_merged, return_indices_ranked_by="self_confidence")

Why isn’t CleanLearning working for me?#

At this time, CleanLearning only works with data formatted as numpy matrices or pd.DataFrames, and with models that are compatible with the sklearn API (check out skorch for Pytorch compatibility and scikeras for Tensorflow/Keras compatibility). You can still use cleanlab with other data formats though! Just separately obtain predicted probabilities (pred_probs) from your model via cross-validation and pass them as inputs.

If CleanLearning is running successfully but not improving predictive accuracy of your model, here are some tips:

  1. Use cleanlab to find label issues in your test data as well (we recommend pooling labels across both training and test data into one input for find_label_issues()). Then manually review and fix label issues identified in the test data to verify accuracy measurements are actually meaningful.

  2. Try different values for filter_by, frac_noise, and min_examples_per_class which can be set via the find_label_issues_kwargs argument in the initialization of CleanLearning().

  3. Try to find a better model (eg. via hyperparameter tuning or changing to another classifier). CleanLearning can find better label issues by leveraging a better model, which allows it to produce better quality training data. This can form a virtuous cycle in which better models -> better issue detection -> better data -> even better models!

  4. Try jointly tuning both model hyperparameters and find_label_issues_kwargs values.

  5. Does your dataset have a junk (or clutter, unknown, other) class? If you have bad data, consider creating one (c.f. Caltech-256).

  6. Consider merging similar/overlapping classes found via cleanlab.dataset.find_overlapping_classes.

Other general tips to improve label error detection performance:

  1. Try creating more restrictive new filters by combining their intersections (e.g. combined_boolean_mask = mask1 & mask2 where mask1 and mask2 are the boolean masks created by running find_label_issues with different values of the filter_by argument).

  2. If your pred_probs are obtained via a neural network, try averaging the pred_probs over the last K epochs of training instead of just using the final pred_probs. Similarly, you can try averaging pred_probs from several models (remember to re-normalize) or using cleanlab.rank.get_label_quality_ensemble_scores.

How can I use different models for data cleaning vs. final training in CleanLearning?#

The code below demonstrates CleanLearning with 2 different classifiers: LogisticRegression() and GradientBoostingClassifier(). A LogisticRegression model is used to detect label issues (via cross-validation run inside CleanLearning) and a GradientBoostingClassifier model is finally trained on a clean subset of the data with issues removed. This can be done with any two classifiers.

from cleanlab.classification import CleanLearning
import numpy as np
from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import GradientBoostingClassifier

# Make example data
data = np.vstack([np.random.random((100, 2)), np.random.random((100, 2)) + 10])
labels = np.array([0] * 100 + [1] * 100)

# Introduce label errors
true_errors = [97, 98, 100, 101, 102, 104]
for idx in true_errors:
    labels[idx] = 1 - labels[idx]

# CleanLearning with 2 different classifiers: one classifier is used to detect label issues
# and a different classifier is subsequently trained on the clean subset of the data.

model_to_find_errors = LogisticRegression()  # this model will be trained many times via cross-validation
model_to_return = GradientBoostingClassifier()  # this model will be trained once on clean subset of data

cl0 = CleanLearning(model_to_find_errors)
issues = cl0.find_label_issues(data, labels)

cl = CleanLearning(model_to_return).fit(data, labels, label_issues=issues)
pred_probs = cl.predict_proba(data)  # predictions from GradientBoostingClassifier

print(cl0.clf)  # will be LogisticRegression()
print(cl.clf)  # will be GradientBoostingClassifier()

How do I hyperparameter tune only the final model trained (and not the one finding label issues) in CleanLearning?#

The code below demonstrates CleanLearning using a GradientBoostingClassifier() with no hyperparameter-tuning to find label issues but with hyperparameter-tuning via RandomizedSearchCV(...) for the final training of this model on the clean subset of the data. This is a useful trick to avoid expensive hyperparameter-tuning for every fold of cross-validation (which is needed to find label issues).

import numpy as np
from cleanlab.classification import CleanLearning
from sklearn.ensemble import GradientBoostingClassifier
from sklearn.model_selection import RandomizedSearchCV

# Make example data
data = np.vstack([np.random.random((100, 2)), np.random.random((100, 2)) + 10])
labels = np.array([0] * 100 + [1] * 100)

# Introduce label errors
true_errors = [97, 98, 100, 101, 102, 104]
for idx in true_errors:
    labels[idx] = 1 - labels[idx]

# CleanLearning with no hyperparameter-tuning during expensive cross-validation to find label issues
# but hyperparameter-tuning for the final training of model on clean subset of the data:

model_to_find_errors = GradientBoostingClassifier()  # this model will be trained many times via cross-validation
model_to_return = RandomizedSearchCV(GradientBoostingClassifier(),
                    param_distributions = {
                        "learning_rate": [0.001, 0.05, 0.1, 0.2, 0.5],
                        "max_depth": [3, 5, 10],
                )   # this model will be trained once on clean subset of data

cl0 = CleanLearning(model_to_find_errors)
issues = cl0.find_label_issues(data, labels)

cl = CleanLearning(model_to_return).fit(data, labels, label_issues=issues)  # CleanLearning for hyperparameter final training
pred_probs = cl.predict_proba(data)  # predictions from hyperparameter-tuned GradientBoostingClassifier

print(cl0.clf)  # will be GradientBoostingClassifier()
print(cl.clf)  # will be RandomizedSearchCV(estimator=GradientBoostingClassifier(),...)
                   param_distributions={'learning_rate': [0.001, 0.05, 0.1, 0.2,
                                        'max_depth': [3, 5, 10]})

Why does regression.learn.CleanLearning take so long?#

To effectively identify errors in a regression dataset, the methods in regression.learn.CleanLearning estimate each datapoint’s aleatoric uncertainty (by fitting a second copy of the regression model to predict the residuals’ magnitudes), as well as its epistemic uncertainty (by fitting multiple copies of the regression model with bootstrap resampling). These uncertainty estimates help provide a robust quality score that accounts for the model’s imperfect predictions.

These uncertainty estimates help produce better results but require longer runtimes. Here are a few options to speed up the runtime of these methods:

  • Reduce the number of bootstrap resampling rounds by decreasing the n_boot argument (default value is 5, set it to 0 to skip the epistemic uncertainty estimation entirely).

  • Set include_aleatoric_uncertainty=False to skip the aleatoric uncertainty estimation.

  • Include less elements in the coarse_search_range argument of regression.learn.CleanLearning.find_label_issues. This is overall set of values initially considered for estimating the fraction of data that have label issues.

  • Reduce the fine_search_size argument of regression.learn.CleanLearning.find_label_issues. A higher number represents a more thorough search to precisely estimate the fraction of data that have label issues.

Below is sample code on how to pass in these arguments.

from cleanlab.regression.learn import CleanLearning

X = np.random.random(size=(30, 3))
coefficients = np.random.uniform(-1, 1, size=3)
y = np.dot(X, coefficients) + np.random.normal(scale=0.2, size=30)

# passing optinal arguments to reduce runtime
cl = CleanLearning(n_boot=1, include_aleatoric_uncertainty=False)
cl.find_label_issues(X, y, coarse_search_range=[0.05, 0.1], fine_search_size=2)

# you can also pass coarse_search_range and fine_search_size as kwargs to CleanLearning.fit
cl.fit(X, y, find_label_issues_kwargs={"coarse_search_range": [0.05, 0.1], "fine_search_size": 2})
CleanLearning(include_aleatoric_uncertainty=False, model=LinearRegression(),
In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

What ML models should I run cleanlab with? How do I fix the issues cleanlab has identified?#

These questions are automatically handled for you in Cleanlab Studio – our platform for no-code data improvement. While this open-source library finds data issues, an interface is needed to efficiently fix these issues in your dataset. Cleanlab Studio is a no-code platform to find and fix problems in real-world ML datasets. Cleanlab Studio automatically runs the data quality algorithms from this library on top of AutoML models fit to your data, and presents detected issues in a smart data editing interface. Think of it like a data cleaning assistant that helps you quickly improve the quality of your data (via AI/automation + streamlined UX). Try it for free!

Stages of modern AI pipeline that can now be automated with Cleanlab Studio

What license is cleanlab open-sourced under?#

AGPL-3.0 license

What does this mean? If you’re working at a company, you can use this open-source library to clean up your internal datasets. You can also use this open-source library to clean up a dataset used to train a model that is deployed in a commercial product. For non-commercial purposes, feel free to release altered versions of the source code as long as you include the same license.

Please email info@cleanlab.ai to discuss licensing needs if you would like to offer a commercial product that utilizes any cleanlab source code.

Can’t find an answer to your question?#

If your question is not addressed in these tutorials, please refer to the: Cleanlab Github issues, Cleanlab Code Examples or our Slack Community.

If your question is not addressed anywhere, please open a new Github issue. Our developers may also provide personalized assistance in our Slack Community.