An AI learning dataset is a collection of examples, often with labels attached, that a machine learning system is trained and tested on, and the quality of the labels and the way the data is split decide how far its reported accuracy can be trusted. A model is normally scored on examples it never saw during training, and a 2021 study of 10 widely used benchmark datasets found label errors in at least 3.3 percent of their test sets on average.
Key takeaways
- Practitioners usually divide labeled data into training, validation, and test sets, because scoring a model on its own training data gives misleadingly high results.
- Information from the test set can leak into a model through repeated tuning or through preprocessing done before the split.
- Northcutt, Athalye, and Mueller found label errors in at least 3.3 percent of test examples on average across 10 benchmark datasets, and at least 6 percent in the ImageNet validation set.
- Datasheets for Datasets, proposed in 2018, ask creators to document motivation, composition, collection process, and recommended uses.
- A dataset description that omits who collected the data, how it was labeled, and what it is not suited for is incomplete.
What a learning dataset contains
Each row of a dataset is an example: an image, a sentence, a row of sensor readings, or a medical record. In supervised learning, every example also carries a label, such as the object in a photo or the correct translation. The system adjusts its internal parameters until its outputs match the labels on the training examples. ImageNet, the benchmark behind much of the deep learning boom, consists of roughly 1.3 million training images sorted into 1,000 classes, and the AlexNet paper of 2012 trained on that set.
Unlabeled data is also used. Large language models, for instance, are trained to predict missing text, so the text itself provides the target. The dataset still shapes the result, because a model can only reflect patterns present in the examples it was given.
Training, validation, and test sets
The scikit-learn documentation, a standard reference for practitioners, states the problem directly: learning the parameters of a prediction function and testing it on the same data is a methodological mistake, since a model that simply repeats the labels it has seen would score perfectly and still fail on new data. This is called overfitting. The remedy is to hold out part of the data as a test set.
A second hold-out is needed because settings such as regularization strength are tuned by trying values and checking the score. If the test set guides those choices, there is still a risk of overfitting to it. The documentation therefore describes a validation set for tuning and a test set kept for final evaluation. An illustrative example: with 10,000 labeled records, a 70/15/15 division gives 7,000 for training, 1,500 for validation, and 1,500 for the final test. The documentation also notes the cost, which is that fewer samples remain for learning, and describes cross-validation as one way to reduce it.
How leakage inflates scores
Leakage happens when information about the test data reaches the model during training. The same scikit-learn guide lists several routes. Preprocessing such as scaling should be fit on the training data only and then applied to held-out data. Shuffling time-ordered data, like news articles, can place near-duplicates of test items in the training set, giving what the documentation calls an inflated validation score. Grouped data, for example several records from one patient, must be kept together so that no group appears on both sides of the split.
| Leak route | What goes wrong | Typical safeguard |
|---|---|---|
| Repeated tuning on the test set | Choices adapt to the test items | Tune on a validation set or by cross-validation |
| Preprocessing before the split | Test statistics shape the training inputs | Fit transformations on training data only |
| Random shuffling of time-ordered data | Test items resemble nearby training items | Split by time |
| Records from one source on both sides | The model recognizes the source, not the pattern | Split by group |
Label errors in the benchmarks themselves
Test sets are assumed to be correct, but they are made by people or by automated pipelines. In a paper presented at NeurIPS 2021, Curtis Northcutt, Anish Athalye, and Jonas Mueller examined 10 commonly used computer vision, language, and audio datasets and estimated that the test sets contained label errors at an average rate of at least 3.3 percent. For the ImageNet validation set the share was at least 6 percent. They also reported that small increases in the share of mislabeled test examples could change which of two models ranks higher, so a leaderboard difference of a point or two may reflect label noise more than model quality. The paper is titled "Pervasive Label Errors in Test Sets Destabilize Machine Learning Benchmarks".
Documenting a dataset
Because data choices shape model behavior, researchers have proposed standard documentation. "Datasheets for Datasets", by Timnit Gebru and six co-authors, first appeared in 2018 and was published in Communications of the ACM in December 2021. Modeled on the datasheets that accompany electronic components, it proposes that every dataset ship with a record of its motivation, composition, collection process, and recommended uses. The aim is better communication between dataset creators and the people who use them. The proposal is voluntary, so the quality of documentation varies from one dataset to the next.
Checklist for reading a dataset claim
- Is the test set separate from the data used for training and tuning, and how was the split made?
- Who labeled the examples, and was label quality measured?
- Does the data match the people, places, and conditions where the system will run?
- Is there a datasheet or similar record of how the data was collected and what it should not be used for?
FAQ
Why can't a model be tested on the data it learned from?
A model can memorize its training examples, so a high score on them says little about new cases. Held-out data estimates how it behaves on examples it has not seen.
Do more examples always make a better dataset?
No. Size helps only when the examples are accurate and match the intended use. Mislabeled or unrepresentative data can be repeated at scale.
What is the difference between a validation set and a test set?
The validation set guides choices made during development. The test set is saved for a final evaluation so that its score is not shaped by those choices.
Is there a legal requirement to document datasets?
Datasheets are a research proposal, not a law. Requirements vary by jurisdiction and sector, and a lawyer or the relevant regulator should be consulted for a specific case.
Related reading
- How training a machine learning model works in practice
- The difference between AI and machine learning
- Deep learning applications and open questions
- A free syllabus for learning AI
More in AI Basics.
Last verified: October 2026
Sources
- Cross-validation: evaluating estimator performance, scikit-learn documentation, 2026
- Pervasive Label Errors in Test Sets Destabilize Machine Learning Benchmarks, Northcutt, Athalye, Mueller, NeurIPS 2021
- Datasheets for Datasets, Gebru et al., arXiv 1803.09010, 2018; Communications of the ACM, 2021
- ImageNet Classification with Deep Convolutional Neural Networks, Krizhevsky, Sutskever, Hinton, NeurIPS 2012