adathalmaz

in English: dataset

basicwidespread

A dataset is a collection of related, organized data—texts, images, numbers—from which artificial intelligence learns or on which its knowledge is measured.

In plain terms

A dataset is to artificial intelligence what a textbook and exercise collection are to a student. The model learns from what is presented to it and only forms an understanding of what appears in the material.

A dataset can consist of anything: millions of photos, medical reports, customer service letters, or a shop's ten years of sales data. The key is that the data is collected for a purpose in a uniform format, and often includes labels, such as "there is a cat in this image".

Different from a pile of files on your computer. A dataset has been curated by someone for a specific purpose and, ideally, cleaned and documented, so its origin and use are known. Its quality often matters more than its size: if it is flawed or biased, the model trained on it will be too.

How it works

Collection: data is gathered from sources such as websites, sensors, corporate systems, or questionnaires.
Cleaning: duplicates, faulty, and incomplete items are filtered out, and everything is brought to a uniform format.
Labeling: if the task requires it, humans or programs mark the correct answer for each example.
Splitting: the dataset is typically cut into three parts. The model learns from the training set, its settings are tuned on the validation set, and it is tested on the test set with examples it has not seen before.
Documentation and maintenance: they describe where the data comes from, what it can and cannot be used for, and update it periodically because the world changes.

An everyday example

When you search for 'dog' in your phone's gallery and your dog photos appear, a model trained on a dataset of millions of labeled photos is at work. It has never seen your photos before; it recognizes them based on learned patterns.

Who is affected

everyoneThe answers from the AI tools you use are determined by the datasets they were trained on, and your public writings and images may also be included in such datasets.

entrepreneur, executiveYour company's own, organized data represents significant value because it is what makes the AI tailored to your business.

developerThe performance of a model or agent often depends more on the quality of the dataset than on the nuances of the code. → Code

lawyer, citizenAssembling datasets raises copyright and data protection issues: whose work is it, whose personal data was included, and under what permission? → Law

A common misconception

MythThe larger the dataset, the better the AI will be.

In realitySize alone is not enough: a smaller but accurate and balanced dataset often yields a better model than a huge but flawed or biased one.

Limits and risks

A dataset is always only a slice of reality: the model knows nothing about what is missing from it, or guesses incorrectly.
If the data is biased—for example, if a group is barely represented—the model will also tend to make unfair decisions.
Datasets become obsolete: a model trained on yesterday's data does not inherently know today's prices, words, and habits.
Many datasets have unclear origins and legal backgrounds, so before using one for corporate purposes, check its license and whether it contains personal data.

Similar, but not the same

database – A database is a queryable, typically continuously updated storage system, while a dataset is a collected, mostly fixed data set assembled for a specific purpose.
training data – The training data is only that part of the dataset from which the model learns; the dataset may also include validation and test portions.
benchmark – A benchmark is a combination of a dataset and measurement rules used to compare models, so it serves for testing, not training.

Related concepts

Broader

adat, machine learning, data processing

Narrower

training data, tesztadat, validation data, synthetic data, benchmark

On the edge

data labeling, bias, data cleaning, fine-tuning, data quality

A brief history

Data collections have long been used in statistics: Ronald Fisher's iris flower data published in 1936 is still taught today. A milestone of the current AI wave was the ImageNet image collection introduced in 2009. The AlexNet model trained on it showed in 2012 that a large, labeled dataset can dramatically improve image recognition.

Where things stand

At the end of 2026, every machine learning-based AI system relies on datasets, and numerous public collections are available on sharing platforms such as Hugging Face or Kaggle. Attention is shifting from sheer volume to quality, transparency of origin, and legality. Since August 2025, the EU AI Act has also required providers of general-purpose models to publish a public summary of their training content.

Status: 2026 November

In the news

Appears in 6 Agent Age stories; the curve shows weekly occurrence in the news magazine.

SK Biopharmaceuticals builds large AI dataset for drug discovery as GaiaBT ties AI to lab validationNov 6, 2026 · Health
Jamendo sues Nvidia again in California over commercial AI training on research music datasetOct 12, 2026 · Jog
Google AI Overviews appear for 80% of branded queries, returning for large brandsSep 30, 2026 · Marketing
Research: more AI agents are reviewing other AI agents' code on GitHubAug 25, 2026 · Code
Nordic AI-Health Initiative outlines federated data infrastructure for medical AIAug 15, 2026 · Health
AI projects aim to strengthen US power grids, from rural Alaska to AWS–PNNL workAug 10, 2026 · Ipar

Also known as

adathalmaz · dataset · dataset · adatszett · dataset · data set · training dataset · training dataset

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.