EU legislation datasets for LLM training: what's available and what it costs

There are three ways to get EU legislation into an LLM training, RAG or eval pipeline: download a free research dataset (MultiEURLEX, EURLEX57K, EUR-Lex-Sum, the EUR-LEX task in LexGLUE, or the eurlex subset of Pile of Law), build your own corpus from EUR-Lex and Cellar, or license a structured corpus such as Arieta Data. The research datasets cost nothing but are static snapshots built for one task, and some carry non-commercial terms. Building your own costs nothing in licence fees, because the EUR-Lex legal notice allows commercial reuse, but costs real engineering time. A licensed corpus costs money (Arieta Data starts at €7,000 a year) and buys structure, history, a citation graph and maintenance: Arieta Data’s 2026-11 snapshot holds 410k+ documents from 16 jurisdictions, split into approx. 4.07M articles.

What counts as an EU legislation dataset?

An EU legislation dataset is a machine-readable collection of the legal acts of the European Union, usually regulations, directives and decisions, with their metadata and, ideally, their text split into recitals and articles. Every EU document has a CELEX number (the GDPR is 32016R0679) and most acts also have an ELI. The authentic text is the one published in the Official Journal of the European Union; the electronic edition has been authentic since 1 July 2013 under Regulation (EU) No 216/2013. Consolidated texts, which merge an act with its amendments, are produced by the Publications Office for information only.

For ML work, the useful properties are not just the text. They are the identifiers that let you join records, the language versions (EU acts exist in up to 24 official languages), the dates (adoption, entry into force, publication), the version history, and the relations between documents: which act amends, repeals or implements which.

Which free research datasets exist?

The best-known public EU legislation datasets were built by academic groups for specific NLP tasks. They are valuable baselines, but each is a one-off snapshot shaped for its task.

Dataset	Source paper	What it contains	Licence as published	Updated?
EURLEX57K	Chalkidis et al., ACL 2019	About 57,000 EU acts with EuroVoc labels, for multi-label classification	See the dataset release	No, static
MultiEURLEX	Chalkidis et al., EMNLP 2021	About 65,000 EU laws in 23 languages with EuroVoc labels	The Hugging Face card says CC BY 4.0 in its licensing section and CC BY-SA 4.0 in its metadata	No, static
EUR-Lex-Sum	Aumiller et al., EMNLP 2022	Up to 1,500 act and summary pairs per language, 24 languages	See the dataset release	No, static
LexGLUE (EUR-LEX task)	Chalkidis et al., ACL 2022	EU legislation classified into 100 EuroVoc concepts, one of seven tasks	CC BY 4.0 per the card	No, static
Pile of Law (eurlex subset)	Henderson et al., NeurIPS 2022	256 GB of mostly US legal and administrative text, with a European laws subset	CC BY-NC-SA 4.0 per the card	No, static

Three limits matter for production use. First, these datasets are frozen at their collection date, so they miss every act and amendment since. Second, they are shaped for classification or summarisation, usually with whole-document text rather than article-level structure. Third, licences vary: Pile of Law is non-commercial and share-alike, and the MultiEURLEX card is internally inconsistent, so a commercial user should read each card and the underlying EUR-Lex terms before relying on it.

If you need aligned language versions for machine translation or multilingual models, the public parallel corpora (DGT-TM, JRC-Acquis) are compared in Parallel corpora of EU law in 24 languages. Filtering a web crawl is another route, with the trade-offs set out in Common Crawl vs curated legal data.

US-centred benchmarks such as LegalBench (162 legal reasoning tasks) are useful for evaluating legal reasoning in general, but they contain little or no EU legislation.

What does it cost to build your own from EUR-Lex and Cellar?

Building your own EU corpus costs nothing in licence fees and a significant amount in engineering. The EUR-Lex legal notice allows reuse of the legal documents on EUR-Lex for commercial or non-commercial purposes with acknowledgement of the source; for Commission documents, Commission Decision 2011/833/EU provides reuse “for commercial or non-commercial purposes”, without charge in principle and without individual application (Articles 4 and 9). CJEU case law is reused under the CURIA copyright notice. The licence question is covered in detail in Is EUR-Lex data licensable for AI training?.

The engineering is where the cost sits:

Task	Why it is harder than it looks
Discovery	Enumerating every act, corrigendum and consolidated version through Cellar’s SPARQL and REST interfaces, and keeping up with new publications
Parsing	Splitting HTML or Formex into recitals, articles, annexes and tables across 24 languages and several decades of layout changes
Versioning	Detecting real text changes and ignoring page-chrome changes, then storing every version with a timestamp
Relations	Turning amendments, repeals, legal bases, corrigenda and national transposition notices into a typed graph
Withdrawals and errors	Tracking corrections and removals, and propagating them to everything downstream
National law	Repeating all of the above per member state, each with its own format, identifiers and licence (the sources per country)

The one thing you cannot build later is an observation history: a record of what each source published, and when, exists only from the day you start observing it. That is the subject of Point-in-time legal data: why version history matters for legal AI evals.

What does a licensed EU legal corpus cost?

A licensed corpus trades a licence fee for work already done: parsing, identifiers, history, the citation graph, licence checks and monthly maintenance. Arieta Data’s indicative pricing is:

Tier	Indicative price	Typical use
Open sample	Free; each record under its upstream licence, Arieta’s structure and metadata under CC BY 4.0; no case law (a CJEU case-law excerpt is available under an evaluation licence on request)	Evaluating fit, academic work, demos
Research / Startup	From €7,000 a year	Internal R&D, RAG, evals and fine-tuning; companies under about 50 staff
Commercial	From €18,000 a year	Legal-AI products in production
Foundation / Lab	Custom, from €50,000	Pre-training, distribution in model weights, custom exports

Commercial and Foundation / Lab licences include 12 months of monthly delta snapshots and a perpetual right to the snapshots delivered during the licence; Research / Startup licences include updates for the term, and the data is deleted on expiry, except records under CC BY and other open licences that permit retention. Renewal for another year of updates costs about 40 to 50 percent of year one. The licence agreement is available on request, with a sample cut for your lanes. Full rights by tier are on the licensing page.

What should you check before choosing a dataset?

Before you train on any legal dataset, check the following. The answers decide whether the data is usable in production and defensible in a model’s documentation.

1. Licence chain per source. EUR-Lex content is reusable under the EUR-Lex legal notice (Commission documents under Decision 2011/833/EU, CJEU case law under the CURIA notice); national sources each have their own terms. One major source, UK case law, needs a separate licence for computational analysis, including machine learning: the Open Justice Licence does not cover it.
2. Personal data. Court decisions name or pseudonymise people. Ask how personal data is handled and how withdrawals reach you.
3. Structure. Article-level records with stable ids save you a parser and make chunking for RAG deterministic; see How to chunk legal documents for RAG.
4. History. Without timestamps per version, you cannot build an eval that respects a cutoff date.
5. Relations. A typed citation graph gives free supervision for retrieval, as described in Citation graphs as supervision signal for legal retrieval.
6. Documentation. A datasheet and Croissant metadata make the dataset loadable and auditable. If you are a general-purpose AI model provider, you also need what goes into your AI Act training-content summary.

Key facts

• Free EU legislation datasets (EURLEX57K, MultiEURLEX, EUR-Lex-Sum, LexGLUE, Pile of Law) are static research snapshots; Pile of Law is CC BY-NC-SA 4.0.
• EUR-Lex content can be reused commercially under the EUR-Lex legal notice with acknowledgement of the source; Commission documents under Decision 2011/833/EU.
• The cost of building your own corpus is engineering: discovery, parsing, versioning, relations, withdrawals and national sources.
• Arieta Data snapshot 2026-11: 419,355 documents, approx. 4.07M articles, approx. 2.05B tokens, 15,098 versions and 801,149 citation edges, from 16 jurisdictions, in 24 languages (EU documents are almost all in English; the national lanes cover 14 national languages).
• Arieta Data pricing starts at €7,000 a year (Research / Startup), €18,000 (Commercial) and €50,000 (Foundation / Lab).

Frequently asked questions

Is there a free EU legislation dataset for commercial use?

Yes, with care. EUR-Lex itself is reusable commercially with attribution, and some research datasets built from it state CC BY 4.0. Check each dataset card: some are non-commercial or share-alike, and the MultiEURLEX card states two different licences.

Why pay for EU legislation when EUR-Lex is free?

The text is free; the work is not. A licensed corpus saves the parsing into articles, the identifiers and dates, the version history, the citation graph, the national sources and the monthly maintenance, and documents the licence chain for each source.

Does Arieta Data include national legislation?

Yes. Snapshot 2026-11 includes national legislation from 15 countries, including France, the UK, Germany, the Netherlands, Spain and Poland: 350k+ national documents in all. National case law is included, delivered with personal data masked.

How often is the data updated?

Arieta Data cuts a snapshot every month. Licensees receive monthly delta snapshots with new and changed records and tombstones for withdrawn ones.

Where Arieta Data fits

Arieta Data is a licensed European legal corpus from Arieta Labs Ltd, the company behind LexAPI. Every snapshot covers EU legislation, CJEU case law and national legislation. Snapshot 2026-11 holds 419,355 documents (61,410 EU documents, of which 22,888 are legislation and 24,180 case law, plus 357,945 national legislation documents from 15 countries), in 24 languages (EU documents are almost all in English; the national lanes cover 14 national languages), with approx. 4.07M articles and approx. 2.05B tokens. National case law is included, delivered with personal data masked.; UK case law is not offered because the Open Justice Licence does not cover computational analysis. Italian Corte costituzionale decisions (CC BY-SA 3.0) are national case law, delivered masked as a separate CC BY-SA file.

The data ships as four products cut from one snapshot: the corpus (article-level text and metadata), the history (every content-hash version observed), the graph (typed citation edges with provenance) and eval sets built to your spec on request. Custom cuts per jurisdiction, language, date range or document type take 5⁠–⁠6 business days. Formats are Apache Parquet and gzip JSON Lines, with a manifest of SHA-256 checksums, a datasheet and per-source attribution. Per-lane counts are on the coverage page. Request a sample: we reply within one business day.

Not legal advice. Check each source’s terms and take your own legal advice before training on any dataset.