Catalog
Datasets cut from one monthly snapshot
Three datasets that share ids, jurisdiction codes and content hashes, plus eval sets built to your spec. Licence the lot, or a custom cut per jurisdiction, language, date range or document type. As of the November 2026 snapshot: the corpus 410k+ documents, the history 15k+ versions, the graph 800k+ edges.
Legal Corpus
Article-level EU and national legal text with full metadata.
id string celex string language string …
410k+documents
Version History
Every content-hash version we have observed, for point-in-time training and evals.
celex string language string jurisdiction string …
15k+versions
Citation Graph
Typed citation edges across EU and national law.
source string target string sourceJurisdiction string …
800k+edges
Legal Eval
Retrieval QA and citation QA sets, built to your spec from the graph and the history.
id string question string goldIds list<string> …
Built to your spec
Compare
Which dataset for which job
| Dataset | What one record is | Use it for | Records |
|---|---|---|---|
| Legal Corpuscorpus | one record per document and language (latest version) | Retrieval-augmented generation over EU and national law | 419,355 |
| Version Historyhistory | one record per document, language and version | Point-in-time evals without future leakage | 15,098 |
| Citation Graphgraph | one record per directed, typed edge | Hard-negative mining and positives for legal retrieval | 801,149 |
| Legal Evaleval | one record per question, built to your spec | Retrieval benchmarks with graph-derived gold sets | to your spec |
Delivery
Same records, two formats, one manifest.
Every SKU is delivered in both formats with identical field names. Each snapshot is immutable and verifiable.
Formats
Apache Parquet
application/vnd.apache.parquet
zstd-compressed, sharded at about 512 MB, one directory per SKU and lane
JSON Lines (gzip)
application/jsonl+gzip
jsonl.gz, same records and field names as Parquet
Files in every snapshot
MANIFEST.json
Record counts, SHA-256 per shard, snapshot time and schema version.
DATASHEET.md
Datasheet for Datasets / Hugging Face style data card for the snapshot.
LICENCES.md
Attribution text per upstream source, generated from the lane manifests.
tombstones/
Ids withdrawn since the previous snapshot; buyers must delete them.
Delivered by signed download or to your bucket. Custom cuts are prepared within 5–6 business days.
Get a sample cut for your use case.
Tell us your jurisdictions, languages and use case. The free sample downloads today; a custom cut by jurisdiction, language, date range or document type, with the licence agreement, is ready within 5–6 business days.
- We reply within one business day
- Custom cuts in 5–6 business days
- Free sample, no sign-up