Glossary
LLM training data explained
LLM training data comes in three kinds, for three stages of building a language model, and each raises different rights questions for sellers.
Updated
Definition
LLM training data is the text, code and human feedback used to build a large language model: broad material for pretraining, worked examples for fine-tuning, and ranked or rated responses for teaching the model which answers people prefer.
Three stages, three kinds of material
People often talk about training data for language models as if it were one pile of text. In practice labs buy and build different material for different stages, and the stage decides what a seller can offer.
| Stage | What the material looks like | What the model gains | What makes it valuable |
|---|---|---|---|
| Pretraining | Very large amounts of general text and code | Grammar, facts, reasoning habits, programming syntax | Breadth, quality and material that is not already public |
| Fine-tuning | Curated examples of a prompt and a good response | How to follow instructions or handle a specialist task | Expert authorship and accuracy |
| Preference data | Several answers to one prompt, ranked or scored by people | Which answers people find helpful or safe | Careful, consistent human judgement |
What sellers can add
Much pretraining text comes from public web archives. Common Crawl, for instance, has been collecting web pages since 2008 and makes its archive available at no charge. Because labs can already reach that material, what sellers can add is usually text that never reached the open web: archives, manuals, transcripts, internal documentation and code.
Worked example: a legal publisher's back catalogue
The following is a made-up scenario for illustration. A small publisher of practice guides for employment lawyers holds twenty years of articles, commentary and annotated forms. An AI company building an assistant for law firms wants writing that explains legal reasoning carefully, which much of the public web does not.
The four steps
- Rights audit: some articles were written by staff, others by freelance contributors. The broker checks the contributor agreements to see whether the publisher holds the rights needed to license them for training. Pieces where it does not are held back.
- Format: the archive is converted from page layouts into clean text with headings kept, because layout debris lowers its value for pretraining.
- Two offers: the full archive is priced as pretraining material; a smaller set of question-and-answer pairs written by the publisher's editors is priced separately as fine-tuning data.
- Terms: the licence allows training and internal evaluation, forbids reproducing articles verbatim in the assistant's answers, and sets out what happens to copies when the term ends.
The two-offer split is the main lesson. The same seller can supply different stages, and buyers value them on different terms.
Copyright and disclosure
Rights are the central question in this market. In the United States, the Copyright Office released a pre-publication version of its report on generative AI training in May 2025. It concluded that whether training is fair use depends on the facts of each case, and it recommended letting voluntary licensing keep developing.
EU disclosure rules
In the EU, the AI Act requires providers of general-purpose models to adopt a policy for complying with EU copyright law and to publish a sufficiently detailed summary of the content used to train them.
What it means for sellers
For a seller, the practical effect is that buyers increasingly want clean provenance: a clear statement of who owns the material and what was licensed, which they can rely on when they disclose their sources.
Mistakes sellers make
- Assuming they own everything on their servers. User posts, licensed photos and contractor work often belong, in part, to someone else.
- Leaving personal data in. Names and contact details in support chats or emails need removing or a lawful basis to share.
- Ignoring duplicates. If half the archive is already on the public web, the buyer may value only the other half.
- Granting unlimited rights. Training, evaluation, redistribution and use in outputs can be licensed separately, and each has a different value.
Sources
3 sources · checked Nov 7, 2026
- Common Crawl: OverviewThat the archive has been gathered since 2008 and is free to use · Checked
- US Copyright Office: Copyright and Artificial IntelligenceThe May 2025 pre-publication report on generative AI training, its fact-specific fair use view and its stance on licensing · Checked
- EU AI Act: Article 53Copyright policy and public training-content summary duties for general-purpose model providers · Checked
LLM training data FAQ
Where do LLMs get their training data?
From a mix of public web archives, books, code repositories, licensed collections bought from publishers and companies, and material written or rated by paid human experts. The mix differs by lab and is often only partly disclosed.
Can I sell data to AI companies for LLM training?
If you own it and it is not already freely available, possibly. Labs look for text, code or expert examples they cannot easily get elsewhere, with clear rights and personal data removed.
What format should LLM training data be in?
Clean text is the usual baseline, with headings and structure kept and page furniture removed. Fine-tuning sets are usually supplied as prompt and response pairs in a simple structured file format the buyer specifies.
Is web-scraped text legal to use for LLM training?
That is being argued in courts and policy reports in several countries. The answer depends on the jurisdiction, the material and how it is used, so licensed material with clear rights is easier for buyers to rely on.
Related
- AI dataset
An AI dataset is the material a model learns from: what it contains, how it is organised, and what buyers check before licensing one.
- Data licensing
Data licensing lets a buyer use a dataset under written conditions while the seller keeps control, which is how most data deals are structured.
- Data ownership
Data ownership is about who can lawfully control, use and license a dataset, which matters far more in a deal than who holds the file.
- Data monetization
Data monetization is the umbrella term for every way an organisation turns its data into money, inside the business or outside it.
- Synthetic data
Synthetic data is generated rather than collected, and labs now use it heavily; this page covers where it replaces real data and where it can't.
Join the DataBroker academy
The DataBroker academy and tools: find companies with data to sell, price it, license it and close the deal.






