Learn AI /AI Foundations /Dataset
Dataset
- Starter
- AI Foundations
- Programme & data teams
- Technical readers
Also called: data set, training data.
Back to the milestone you came from
In five seconds
What it is
A dataset is an organised collection of examples that a system can use for analysis, training, testing or search.
In plain words
A dataset is a prepared collection of examples used for a task, such as training a model, checking how well it works, or building search.
An everyday example
A pile of receipts becomes a dataset when each purchase is entered in a consistent table, with one row per receipt and columns for the date, shop, amount and spending category. The organised records can then be searched, compared or totalled. Missing receipts, duplicate entries and inconsistent categories will affect the results, just as gaps and errors affect a technical dataset.
A real dataset is often much larger and more varied than a neat folder of checked sentence pairs. Many datasets are not manually reviewed line by line, and some have no labels at all. Also, a translation system does not simply look up identical past sentences; in machine learning it uses many examples to adjust a model statistically, so the folder analogy can make the process seem more direct and exact than it really is.
Remember thisA dataset is the organised collection of records used for analysis, training, testing or search. It is an input to a system, not the model itself.
Understand it
Understand it
Why it matters
If someone says an AI system was trained, tested or evaluated, the next practical question is: on what dataset? The quality, coverage and limits of that dataset strongly affect what the system can do, where it may fail, and who may be left out. A small, outdated or biased dataset can produce misleading results even if the model is advanced. For non-specialists, understanding datasets helps with procurement, risk review, quality checks, and basic questions about fairness, privacy, documentation and fitness for purpose.
How a dataset is prepared and used
After viewing this, the reader should be able to explain that a dataset is a structured collection of examples that is prepared, sometimes split, and then used to produce something such as a model, score or search index.
01
1. Collect examples
Start with items relevant to the task, such as documents, images, audio clips or sentence pairs.
Draw a stack of mixed source items on the left with small icons for document, image and audio, feeding into the next stage with an arrow.
02
2. Organise the collection
Put the items into a consistent structure, such as rows in a table, files in folders, or documents with labels and metadata.
Show the mixed items becoming a tidy grid or labelled file set. Add small tags like date, language or category.
03
3. Clean and split if needed
Fix errors, remove duplicates, and optionally separate some examples for validation or testing.
Show a filter icon removing duplicate and error symbols, then split the cleaned block into a larger section and a smaller held-out section.
04
4. Use it for a task
The prepared dataset can be used to train a model, evaluate performance, build a search index, or support analysis.
From the organised and split data, draw arrows to three outputs on the right: model, score chart and search index.
A left-to-right flow diagram with four stages: collected examples, cleaned and organised dataset, optional split into training and test portions, and outputs such as a trained model or search index. Read from left to right to see what goes in, what is prepared, and what comes out.
Hover or tap each stage to reveal a short caption with examples of what is added, changed or produced at that stage.
Describe this without the diagram
This diagram explains a dataset as part of a process. First, examples are collected, such as documents, images, audio clips or sentence pairs. Second, the examples are organised into a consistent structure, for example a table or labelled file collection. Third, the collection is cleaned by fixing errors and removing duplicates, and it may be split into one part for building a system and another part for checking performance. Fourth, the prepared dataset is used for a task, producing outputs such as a trained model, an evaluation score or a search index. The key idea is that a dataset is an organised collection of examples, not the final system.
With reduced motion: Show the same four stages as a static left-to-right flow with labelled arrows and no animation: collected examples, organised dataset, cleaned and optional split, then outputs such as model, score and search index.
A dataset is a grouped collection of information arranged so people or software can use it. It might be a table of rows and columns, a folder of images, a set of audio clips, or many documents with labels. Each item in the collection is usually one example.
What goes in depends on the task. For language work, a dataset might contain sentences and their translations. For image work, it might contain pictures and a label such as “road” or “tree”. For search, it might contain reports, guidance notes and keywords. People often clean the dataset first, which means fixing errors, removing duplicates, and making the format consistent.
What happens next also depends on the use. A machine-learning system may use the examples to find statistical patterns. A search tool may index the documents so they can be found quickly. A test set is a separate part of the data used to check performance rather than build the system.
What comes out is not the dataset itself but a result produced from it, such as a trained model, a benchmark score, a search index, or an analysis. A dataset is not the same as data in general: “data” is the broad category, while a dataset is one defined collection prepared for a purpose. It is also not the same as a model, which is a mathematical pattern adjusted using examples during training.
How “Turning a pile of receipts into a table” maps onto the real thing
| In the story | In the AI system |
|---|---|
| Shared folder of past sentence pairs | Dataset containing training examples |
| Each source sentence matched with a checked translation | One labelled example |
| Removing repeats and fixing formatting | Data cleaning |
| Setting aside a small portion for later checking | Validation or test set |
A worked example
An officer in an evaluation office needs a classifier to tag recommendations from published evaluation reports before an annual progress report goes for clearance.
Starting from: 1,200 recommendation extracts, each reviewed and labelled as governance, programme design or operations.
- The officer removes 90 duplicate extracts and 30 records with missing text or labels.
- The remaining records are stored in the same format, with one recommendation and one reviewed label per record.
- The 1,080 records are divided into 860 for training, 110 for validation and 110 for testing.
- The training set is used to fit the classifier, the validation set to choose settings, and the untouched test set for the final evaluation.
Result: The dataset is the organised collection of 1,080 labelled recommendation extracts. Its three separate portions allow the classifier to be fitted, selected and evaluated without using the same records for every step.
In a United Nations workplace
The section chief in a lightly staffed field presence needs a situation report cleared by close of business, while several documents are still awaiting translation and the connection is unreliable. Previously cleared source text and checked translations are organised as paired records on approved local infrastructure.
Those paired records form a dataset. Each may include the source sentence, its checked translation, language and document type. The collection can support translation suggestions or search for earlier wording, but it is not itself a translation model.
Before you try thisThe collection must be limited to material cleared for this use. Staff must not add names, contact details, witness accounts, protection information or security-sensitive passages to any tool not approved for such data. Disclosure could expose individuals or operations, and suggested translations still require human review.
Where you may hear it
- A vendor demo to an ICT governance board
- A procurement meeting reviewing data requirements
- An evaluation office discussion of test results
- A country office session on document search
Avoid confusion
Avoid confusion
Often confused with
Data
Recorded observations, measurements or examples that people or systems can store, examine and use for a purpose.
Corpus
An organised collection of text or recorded speech used to study language or train and test language-processing systems.
Training
Training is the process of adjusting a model's parameters using data or feedback to improve performance on a learning objective.
Model
A parameterised mathematical function learned from data that maps inputs to outputs such as labels, scores, predictions or generated text.
Common misconceptions
Common misconception: A dataset is just any pile of information.
Correction: Not quite. A dataset is usually a defined collection organised for a purpose, with some structure that lets people or software use it consistently. Loose notes scattered across systems may be data, but not yet a workable dataset.
Common misconception: Training data and dataset always mean exactly the same thing.
Correction: People often use them loosely, but they are not always identical. A dataset may be used for training, validation, testing, search or analysis. Training data is specifically the part used to adjust a model during training.
Common misconception: A bigger dataset always gives better results.
Correction: Size helps only if the examples are relevant, accurate and diverse enough for the task. A very large but noisy, biased or outdated dataset can still produce poor or harmful results.
Common misconception: If a dataset comes from the real world, it must be neutral.
Correction: Real-world collections reflect how they were gathered, labelled and filtered. Missing groups, historical inequalities and measurement errors can all shape the dataset before any model is built.
Quick check
Which statement best describes a dataset in AI work?
Option 2 is best because a dataset is a defined collection used for a task. Option 1 describes only training data, not every dataset. Option 3 describes a schema without records. Option 4 describes a model learned from examples, not the dataset.
Go deeper
Go deeper
The simple version
A dataset is a prepared collection of examples used for a task, such as training a model, checking how well it works, or building search.
The technical version
In technical use, a dataset is a finite collection of observations or records, usually stored in a defined schema or file structure, that can be processed by software. Each record may contain features, labels, metadata or raw content such as text, images, audio or logs. In machine learning, a dataset is often divided into training, validation and test sets so that parameter fitting, model selection and final evaluation are kept separate. The exact form varies by task: a tabular dataset has rows and columns, a text corpus contains documents or sequences, and a multimodal dataset may link text with images or audio. The term is broader than “training data”, because the same overall dataset may include held-out subsets for evaluation or may be used for indexing, retrieval or descriptive analysis rather than training.
Reviewed external resources
Primary sourceRecommendation on the Ethics of Artificial Intelligence (opens in a new tab)UNESCO
This link gives governance and ethics context for how data should be handled in AI systems and suits policy, programme and oversight readers.
Intermediatelink checked 2026-07-31
EncyclopaediaData set (opens in a new tab)Wikipedia
This link gives a broad introductory overview of the term and suits readers who want a quick general reference.
Intermediatelink checked 2026-07-31
ArticleMachine Learning Glossary (opens in a new tab)Google
This link gives a practical glossary-style definition in machine-learning language and suits readers who want concise technical wording.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Dataset (opens in a new tab)YouTube search
This link is a search results page for video explanations of the term, useful for readers who prefer to compare several visual or spoken introductions.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Dataset (opens in a new tab)Reddit search
This link is a search results page for community discussions of the term, useful for readers who want to see common questions, informal explanations and points of confusion.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
A dataset is the organised collection of records used for analysis, training, testing or search. It is an input to a system, not the model itself.