← All articles
Blog

What Is an AI Dataset? Types, Structure and Sources

An AI dataset is a structured collection of examples used to train, fine-tune, evaluate or ground an artificial intelligence system. It might be millions of web pages for pre-training a language model, a few thousand labelled support tickets for a classifier, or a carefully built set of business cases used to test an agent. This guide explains what an AI dataset is, the main types, what makes one useful and where teams source real-world data.

The parts of an AI dataset

Most AI datasets contain:

  • Examples: the inputs, such as text, images, audio, documents or interaction logs.
  • Labels or targets: the desired outputs or known outcomes, where relevant. See what is data annotation.
  • Metadata: source, date, language, format, licence and collection method.
  • Splits: separate training, validation and test sets.
  • Documentation: a datasheet or data card describing contents, intended use, limitations and provenance.

Types of AI dataset

Pre-training datasets

Very large, broad collections of text, code, images or audio used to give a model general knowledge and skills.

Fine-tuning datasets

Smaller, targeted sets used to adapt a model to a task, domain or style, including instruction and preference data.

Evaluation datasets

Benchmarks and test sets with known answers, used to measure performance. They must be kept separate from training data to avoid contamination.

Retrieval datasets

Document collections indexed so a model can look up information at run time, as in retrieval-augmented generation.

Environment and trajectory datasets

Records of multi-step tasks, including states, actions and outcomes, used to train and test agents with reinforcement learning. See RL environment data.

What makes an AI dataset useful

Relevance

The data should reflect the task and the domain where the model will be used.

Quality

Accurate labels, little noise, consistent formats.

Coverage

Enough variety to include edge cases, exceptions and failures, not just typical examples.

Linkage

For workflow tasks, related records should be connected: the request, the correspondence, the change and the outcome.

Temporal integrity

Timestamps and versioning that allow a model to be tested on what was known at a given moment.

Clear rights and provenance

Documented sources and a licence that covers the intended use. See data provenance.

Privacy

Personal data removed or protected appropriately.

Where AI datasets come from

  • Public and open data: web crawls, open government data and open-licensed collections.
  • Licensed content: publishers, media libraries and data vendors.
  • Human-created data: annotation and expert-written examples.
  • Synthetic data: generated by models or simulations. See synthetic vs real data.
  • Operational business records: licensed from companies that created them in the course of real work.

Our guide on where to buy AI training data compares these routes.

Why business records are a distinct kind of AI dataset

Public data shows what people publish. Business records show how work is done. A quotation, its revisions, the purchase order, the production record and the invoice together capture decisions under real constraints. That makes them well suited to evaluating assistants and agents that will work inside companies, and to building realistic environments.

They also bring challenges: mixed formats, missing links, personal data and confidentiality. Turning them into a usable AI dataset means defining the task, finding suitable sources, establishing rights, and then cleaning and linking the records.

Documenting a dataset

Good practice is to publish or keep a datasheet covering:

  • motivation and intended uses
  • composition and collection process
  • preprocessing, cleaning and labelling
  • known limitations and biases
  • licence, provenance and restrictions
  • maintenance and versioning

How Maevar helps

Maevar introduces record owners to vetted buyers, typically AI teams looking for business records for training, evaluation and retrieval. We start with your task. The buyer pays our introduction fee. The owner and the buyer agree terms, preparation and how personal details are handled, directly. Discuss your use case or read how to write a data licensing agreement.

Common mistakes when building an AI dataset

  • Mixing evaluation and training data: even a small overlap inflates scores.
  • Collecting only typical cases: models then fail on the exceptions that matter in production.
  • Dropping metadata: without dates, versions and sources, it becomes impossible to test temporal reasoning or trace problems.
  • Breaking links during cleaning: removing identifiers can disconnect related records.
  • Ignoring licence scope: data licensed for evaluation may not be licensed for training.
  • No documentation: undocumented datasets are hard to reuse, audit or defend.

Versioning and maintenance

Datasets change as errors are fixed, sources are added and guidelines evolve. Give every version an identifier, keep a changelog, and record which model was trained or evaluated on which version. Treat a dataset as a maintained product rather than a one-off file.

Choosing formats

Keep original files where layout or media matter, such as PDFs, drawings, audio and images, alongside structured extracts. For linked business data, a relational or graph structure with stable IDs often preserves the relationships better than flat files.

FAQ

What is an AI dataset in simple terms?

A collection of examples, often with answers, used to teach or test an AI system.

How big does an AI dataset need to be?

It depends on the purpose. Pre-training needs vast volumes, while evaluation sets can be small if they are well designed.

What is the difference between training and test data?

Training data teaches the model. Test data measures it, and must be kept separate.

Can I use any data I find online?

Not necessarily. Copyright, terms of service and privacy law may restrict use. Check provenance and licences.

What format should a dataset be in?

Whatever suits the task. JSONL, Parquet and CSV are common for text and structured data, with original files kept for documents and media.

Maevar privacy notice

21 October 2026

Who we are

Maevar is operated by Marcias Limited, company number SC653353. Marcias Limited is responsible for the personal information covered by this notice.

Registered office: 51 Watt Road, Hillington, Glasgow, G52 4RY, United Kingdom.

For privacy questions or requests, email hello@maevar.com.

This notice covers this website and enquiries sent to Maevar. Any proposed introduction of a record owner to a buyer, or any sale or licence they agree, requires a separate review of the relevant rights, privacy requirements and agreed terms.

Enquiries and information you provide

The enquiry fields help you prepare an email containing your work email address, organisation, task summary and any optional format or timing details. The draft is prepared on your device. The website does not submit or store it: you choose whether to open, review and send it using your email app. There is no file-upload facility.

Please provide only a non-confidential summary. Do not include datasets, personal records, passwords or other sensitive information.

When you send an email, we receive the message, your contact details and any other information you choose to include. Richard McKay handles enquiries. Authorised service providers and connected tools, including AI-assisted tools, may support the handling of correspondence.

Why we use your information

We use enquiry information to respond, understand your requirements, assess whether we can help and discuss relevant business opportunities. We rely on our legitimate interests in managing business enquiries and developing appropriate business relationships, while considering your rights and reasonable expectations.

We also have a legitimate interest in keeping the website and email services secure and reliable. Where we must retain or disclose information to meet a legal obligation, we rely on that obligation.

Website and email services

We use OpenAI's ChatGPT Sites to host this website, Cloudflare for website delivery, security and email routing, and Google Workspace to receive and manage email enquiries. These services process information needed to deliver and protect the website and email service. This can include IP addresses, browser information, access times and technical or security logs, as well as email information when you contact us.

Optional analytics is controlled separately from necessary hosting and security services. Cloudflare uses a bot-protection cookie called __cf_bm, which normally expires after 30 minutes of inactivity. More information is available in Cloudflare's cookie documentation.

Optional analytics and your choice

Google Analytics loads only after you accept analytics. Rejecting keeps optional analytics off.

We load Google Analytics only after you accept it. Rejecting does not stop you using this website or contacting us. Necessary hosting, delivery and security services continue to operate, including their technical logs.

With your permission, analytics measures page visits and limited website interactions. We use fixed page information without search queries or URL fragments. We do not send your name, email address, organisation, enquiry text, form answers or the contents of email-draft links to Analytics. Analytics is not a record of the contents of your enquiry.

The email preparation tool may record that an email draft was prepared, without any of its contents. This is an expression of interest only: the website cannot tell whether you send the email. It is not counted as a submitted enquiry.

No Google Ads, Meta or LinkedIn advertising tags are activated by these controls. Advertising storage, advertising user data and ad personalisation remain disabled. We do not use enhanced conversions or upload contact details to Analytics.

Your privacy choice is stored on this website in your browser for up to six months. If you accept analytics, Google Analytics may set first-party cookies named _ga and _ga_ followed by the property identifier. Our code sets a maximum cookie lifetime of six months; your browser may apply a shorter limit. Google processes the analytics information and may process it outside the UK. See Google’s privacy policy. Our Google Analytics property retains user-level and event-level data for two months, without restarting that retention period when you return. This setting does not limit how long Google retains standard aggregated reports. Optional Google data sharing, Google Signals, user-provided data collection, granular location and device collection, and advertising personalisation are switched off.

Use Cookie settings in the footer to change your choice. Withdrawing analytics consent stops new measurement and reloads the page into a no-analytics state, so save any unfinished enquiry first. It also removes the relevant first-party analytics cookies where the browser permits. This does not erase information already sent to Google. We do not replay interactions that happened before you consented.

Our service providers may process information outside the UK. Their published data-processing terms describe the applicable international-transfer safeguards, including recognised adequacy arrangements and contractual safeguards:

Contact hello@maevar.com if you would like further information about these arrangements.

How long we keep information

We review unsuccessful enquiries 12 months after the last substantive contact and delete them unless there is an ongoing business or legal reason to retain them. For example, information may still be needed for an active discussion, to meet a legal obligation, or to establish, exercise or defend legal claims. Where we retain information longer, we keep it only for as long as that reason applies.

Enquiries that lead to a business relationship and related contract records are kept according to the needs of that relationship and applicable legal requirements. The 12-month enquiry review period is not a guarantee that every provider's security log or backup is erased on the same date; those records are subject to the relevant service's retention and deletion arrangements.

Your rights and complaints

Depending on the circumstances, you may ask to access, correct or delete your personal information, restrict its use, or receive a portable copy. You may also object to processing based on legitimate interests. These rights are subject to applicable conditions and exceptions.

Email hello@maevar.com to make a request or raise a concern. You can also complain to the Information Commissioner's Office through ico.org.uk/make-a-complaint.

Changes to this notice

We will update this notice when our website or information-handling arrangements change. The date above identifies the version you are reading.