davidwickerhf/cjeu-opendata

CJEU / CELLAR Case Law

A full-coverage corpus of European Court of Justice case law — every judgment, order, opinion, and notice the Court of Justice and General Court have published since 1954, plus the national-court decisions that cite EU law (sector 8). Each case carries its own metadata, multi-language full text where available, and the inbound + outbound citation graph as joinable CELEX identifiers.

The dataset is a single source of truth for empirical EU-law research: who cites whom, what subject matters dominate, how the procedural-language mix has shifted over seven decades, where the General Court's IP docket fits in.

Table	Granularity	Description
cases.parquet	one row per ECLI	All case metadata. 107 columns covering court formation, judicial procedure type, advocate general, judge rapporteur, subject matter, eurovoc concepts, citing / cited-by edges, etc. Per-field docs in FIELDS.md.
fulltexts.parquet	one row per ECLI	Plain-text body of each document plus provenance flags (text_source, text_format, text_language, missing_reasons).

Headline numbers

• Date range: 1954-11-10 → 2026-05-21
• Cases: 46,180
• Fulltexts: 546,733
• Last refreshed: 2026-05-29T09:23:55+00:00

Coverage

Per-decade fulltext availability

Decade	Cases	With fulltext	Coverage
1950s	84	84	100.0%
1960s	555	555	100.0%
1970s	1,683	1,683	100.0%
1980s	3,784	3,783	100.0%
1990s	5,696	5,653	99.2%
2000s	8,939	8,932	99.9%
2010s	15,076	14,899	98.8%
2020s	10,362	10,315	99.5%

In total, 542,934 of 546,733 (99.3%) documents have a non-trivial body text (≥200 characters). Rows without body text carry a populated missing_reasons column:

• SUMMARY_UNAVAILABLE_UPSTREAM — 90,829 rows
• FULLTEXT_UNAVAILABLE_UPSTREAM — 3,414 rows
• UNAVAILABLE_UPSTREAM — 192 rows

Sector split

Sector	Description	Cases	Share
6	EU courts (CJEU / GC / CST)	44,375	96.1%
8	National case law citing EU law	1,817	3.9%

Languages

Each document is available in its original procedural language and, where CELLAR provides them, in additional EU-official-language translations. Top 10 by row count in fulltexts.parquet:

FR (41,157), EN (34,195), DE (34,069), IT (32,782), NL (32,374), DA (31,240), EL (30,359), PT (29,579), ES (28,494), SV (24,642)

Notes

• Schema fields that apply only to legislation (e.g. eli, in_force) or that describe the body text rather than the case (text_language, text_format, text_source) are not present in cases.parquet. See FIELDS.md for the full per-field reference.
• Each row in both tables carries a __source_window column identifying the scrape window(s) it came from.

Quick start

Pandas (recommended for analytics)

import pandas as pd

# Direct parquet — fastest, no datasets dep required.
URL = "https://huggingface.co/datasets/davidwickerhf/cjeu-opendata/resolve/main"
cases = pd.read_parquet(f"{URL}/cases.parquet")
texts = pd.read_parquet(f"{URL}/fulltexts.parquet")

print(cases.shape, texts.shape)

HuggingFace datasets (streaming-friendly)

from datasets import load_dataset

cases = load_dataset("davidwickerhf/cjeu-opendata", "cases", split="train")
texts = load_dataset("davidwickerhf/cjeu-opendata", "fulltexts", split="train")

# Streaming for the fulltexts table (avoids loading 325 MB into RAM):
texts_stream = load_dataset(
    "davidwickerhf/cjeu-opendata", "fulltexts", split="train", streaming=True
)
for row in texts_stream:
    print(row["ecli"], len(row["text"]))
    break

Polars (fast column scans)

import polars as pl

cases = pl.read_parquet(
    "https://huggingface.co/datasets/davidwickerhf/cjeu-opendata/resolve/main/cases.parquet"
)
print(cases.select(["ecli", "celex", "subject_matter"]).head())

Recipes

Practical queries you can paste verbatim. All examples assume cases and texts DataFrames loaded as in Quick start.

1. Filter by date range and sector

import pandas as pd

cases["pub_date"] = pd.to_datetime(
    cases["date_publication"].str.split(";").str[0],
    errors="coerce", utc=True,
)
recent_cjeu = cases[
    (cases["pub_date"] >= "2020-01-01")
    & (cases["sector"] == "6")
]
print(len(recent_cjeu), "post-2020 sector-6 cases")

2. Subject-matter filter (preliminary references on VAT)

vat_refs = cases[
    cases["subject_matter"].fillna("").str.contains("Value added tax")
    & cases["type_procedure"].fillna("").str.contains(
        "Reference for a preliminary ruling"
    )
]
print(f"{len(vat_refs):,} VAT preliminary references")
print(vat_refs[["ecli", "origin_country", "date_publication"]].head())

3. Build a citation network (in-dataset edges only)

import pandas as pd

# work_cites_work is CELEX-form; explode for one edge per row.
edges = (
    cases[["celex", "work_cites_work"]]
    .assign(target=lambda d: d["work_cites_work"].str.split(";"))
    .explode("target")
    .dropna(subset=["target"])
    .query("target != ''")
    .rename(columns={"celex": "source"})
    [["source", "target"]]
)
# Restrict to edges whose target is in the dataset (self-joinable):
in_dataset = set(cases["celex"].dropna())
edges = edges[edges["target"].isin(in_dataset)]
print(f"{len(edges):,} case→case citation edges")

4. PageRank over the citation graph (with NetworkX)

import networkx as nx

g = nx.from_pandas_edgelist(
    edges, source="source", target="target", create_using=nx.DiGraph,
)
pr = nx.pagerank(g, alpha=0.85)
top = sorted(pr.items(), key=lambda kv: -kv[1])[:10]
for celex, score in top:
    row = cases.loc[cases["celex"] == celex].iloc[0]
    print(f"{row['ecli']:30s}  PR={score:.4f}  {row['subject_matter'][:60]}")

5. Fetch the body text for a specific case

# texts has one row per (celex, language). Modern cases publish in ~24
# languages; you usually want the procedural language (matches `language_procedure`).
def fetch_text(ecli: str, lang: str | None = None) -> str | None:
    rows = texts[texts["ecli"] == ecli]
    if rows.empty:
        return None
    if lang:
        rows = rows[rows["text_language"].str.upper() == lang.upper()]
    if rows.empty:
        return None
    return rows.iloc[0]["text"]

body = fetch_text("ECLI:EU:C:2014:317")  # the Google Spain case
print(body[:500] if body else "(no fulltext available)")

6. Per-language fulltext counts

print(
    texts[texts["text"].str.len() >= 200]
    ["text_language"].value_counts().head(15)
)

7. Cases without fulltext, with the reason recorded

no_text = texts[texts["text"].str.len() < 200]
print(no_text["missing_reasons"].value_counts().head())

8. Join cases and texts on ECLI

joined = cases.merge(texts, on="ecli", how="left", suffixes=("", "_text"))
print(joined.shape)  # ~46 k rows (one per ECLI; no fan-out post-dedup)

Citation graph

Each case row carries two citation columns, both as ;-separated lists of CELEX identifiers:

• work_cites_work — outbound edges (cases / acts this case cites)
• cited_by — inbound edges (cases that cite this case)

Topology

• Total outbound edges: 708,913
• Inside the dataset (case → case, self-joinable): 466,092 (65.7%)
• Outside the dataset (case → legislation / treaty / opinion): 242,821

Most-cited cases (top 10 by inbound count)

Rank	ECLI	Cited by	Year	Subject matter
1	ECLI:EU:T:2003:199	1,491	2003	Trade marks;Intellectual, industrial and commercial property
2	ECLI:EU:C:1998:442	1,343	1998	Intellectual, industrial and commercial property;Approximati…
3	ECLI:EU:C:2007:333	1,322	2007	Trade marks;Intellectual, industrial and commercial property
4	ECLI:EU:T:2006:397	1,020	2006	Trade marks;Intellectual, industrial and commercial property
5	ECLI:EU:T:2009:14	989	2009	Trade marks;Intellectual, industrial and commercial property
6	ECLI:EU:T:2007:46	862	2007	Trade marks;Intellectual, industrial and commercial property
7	ECLI:EU:C:1998:154	800	1998	State aids;Competition
8	ECLI:EU:C:1999:323	773	1999	Intellectual, industrial and commercial property;Approximati…
9	ECLI:EU:C:2014:2258	727	2014	Agriculture and Fisheries;External relations;Wine
10	ECLI:EU:C:2005:408	652	2005	Competition;Agreements, decisions and concerted practices;Co…

What's in the corpus

Top subject-matter atoms

subject_matter is a ;-separated list of EU legal-area atoms. The single-atom counts below are after exploding the lists.

Subject atom	Cases
Intellectual, industrial and commercial property	5,806
Staff regulations and employment conditions - EC	5,280
Trade marks	4,729
Competition	4,711
Agriculture and Fisheries	4,622
Approximation of laws	4,591
Free movement of goods	2,821
Taxation	2,704
External relations	2,591
Right of establishment	2,572
Freedom to provide services	2,362
Judicial cooperation in civil matters	2,150
State aids	2,105
Agreements, decisions and concerted practices	2,029
Environment	1,931

Top procedure types

Procedure	Cases
Reference for a preliminary ruling	16,825
Action for annulment - unfounded	7,551
Action for annulment	4,634
Action for failure to fulfil obligations - successful	3,370
Appeal - unfounded	3,042
Action for annulment - successful	2,840
Action for annulment - inadmissible	2,286
Staff cases - unfounded	2,203
Appeal - inadmissible	1,788
Action for damages - unfounded	1,497

Top origin countries

origin_country records the member state that referred the case (preliminary references) or whose national court the action originated from. Counts after exploding multi-country cells.

Country	Cases
Germany	8,374
Italy	4,414
France	3,806
Netherlands	3,009
Belgium	2,973
Spain	2,605
United Kingdom	2,322
Austria	1,976
Greece	1,198
Poland	1,006
Portugal	929
Ireland	736
Luxembourg	700
Denmark	633
Romania	596

Working with fulltexts

fulltexts.parquet has 8 columns:

Column	Type	Description
ecli	string	Join key against cases.parquet.
celex	string	Sometimes multi-valued (62019CJ0793;62019CJ0793_RES) when an ECLI bundles multiple work items.
text	string	Plain text. Empty when CELLAR has no body for this work.
text_source	string	CELLAR_ITEM / CELLAR_REST_XHTML / INFOCURIA_BLOB_HTML / EXTRACTOR_FALLBACK_TEXT.
text_format	string	html / xhtml / pdf / xml. Original markup format before plain-text extraction.
text_language	string	ISO 639-1 code (FR, EN, DE, …). The procedural language at the CJEU is historically French, hence FR dominance.
missing_reasons	string	;-separated tags explaining empty fields, e.g. FULLTEXT_UNAVAILABLE_UPSTREAM.
__source_window	string	Internal: the date window(s) that scraped this row. ;-joined post-dedup.

The body is plain text — no markup, no headers/footers, no page numbers. For semantic analysis (sentence segmentation, embeddings) just iterate over text directly:

from itertools import islice

def iter_long_judgments(min_chars: int = 5000):
    long_ones = texts[texts["text"].str.len() >= min_chars]
    for _, row in long_ones.iterrows():
        yield row["ecli"], row["text_language"], row["text"]

for ecli, lang, body in islice(iter_long_judgments(), 3):
    print(f"--- {ecli} ({lang}, {len(body):,} chars) ---")
    print(body[:300], "…\n")

How the data was extracted

The pipeline lives in two repos:

1. cellar-extractor — the actual scraper. Hits the CELLAR SPARQL endpoint for metadata + citation graph, plus InfoCuria + CELLAR REST for body text and provenance flags. Handles per-CDM-predicate flattening, sector-3 legislation support, citation URI→CELEX resolution.

2. cjeu-migration — the orchestrator. Iterates date windows (month-sized), retries failed windows with exponential backoff, persists per-window CSV + JSON checkpoints (survives crashes), then consolidates everything into the two parquet files you're reading here. Includes the cleanup scripts (scripts/cleanup_hf_dataset.py) that produced this version.

Single-command reproduction:

git clone https://github.com/davidwickerhf/cjeu-migration
cd cjeu-migration
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
cp .env.example .env  # set HUGGINGFACE_TOKEN if you intend to upload
cjeu-migrate run --skip-upload  # local-only; drop --skip-upload to push to HF

The full corpus run (1954 → today) takes ~4-6 h on a 16-thread CPU instance.

Field provenance

See FIELDS.md for per-field source documentation: which CDM predicate or InfoCuria key produced each column, type, cardinality, and whether it's case-law-only or any-sector.

Schema

Canonical columns

These are the contracted columns — they appear on every row (populated or null), and their semantics are stable across runs:

• advocate_general
• affecting_ids
• affecting_string
• alternate_identifiers
• based_on_treaty
• celex
• citations_extra_info
• cited_by
• citing
• commented_by_agent
• conclusions
• creation_date
• date_of_creation
• date_of_request
• date_publication
• delivered_by_court_formation
• directory_codes
• ecli
• eurovoc
• fulltext_source
• internal_status_code
• judge_rapporteur
• judicial_procedure_type
• keywords
• language_procedure
• legal_resource
• local_identifier
• missing_reasons
• national_judgement
• natural_number_celex
• origin_country
• origin_country_or_role_qualifier
• references_journals
• resource_legal_type
• resource_type
• sector
• subject_matter
• summary
• summary_source
• type_procedure
• work_date_creation_legacy
• work_version
• year_of_resource

Discovered columns

CDM predicates surfaced opportunistically from CELLAR that aren't part of the canonical contract but were populated for at least one row in this snapshot:

• case_law_affaire_jurisdiction
• case_law_affaire_number
• case_law_affaire_type
• case_law_affaire_year
• case_law_amends_resource_legal
• case_law_confirms_resource_legal
• case_law_declares_incidentally_valid_resource_legal
• case_law_declares_valid_resource_legal
• case_law_declares_void_by_preliminary_ruling_resource_legal
• case_law_declares_void_resource_legal
• case_law_defended_by_agent
• case_law_delivered_by_advocate_general
• case_law_delivered_by_court_national
• case_law_delivered_by_judge
• case_law_immediately_enforces_resource_legal
• case_law_incidentally_declares_void_resource_legal
• case_law_interpretes_judgement_resource_legal
• case_law_is_about_case_law_subject_matter
• case_law_is_about_concept_case_law
• case_law_is_about_concept_new_case_law
• case_law_joins_case_court
• case_law_national_act_reference_european
• case_law_national_act_reference_international
• case_law_national_act_reference_national
• case_law_national_based_on_resource_legal
• case_law_national_decision_internal_identifier
• case_law_national_follow_up
• case_law_national_judgement_reference
• case_law_national_keywords
• case_law_national_parties
• case_law_national_reference_publication
• case_law_national_reference_publication_conclusion
• case_law_published_in_erecueil
• case_law_reexamined_by_case_court
• case_law_referred_to_for_preliminary_ruling_case_law
• case_law_requested_by_agent
• case_law_states_failure_concerning_resource_legal
• case_law_subject_to_appeal_in_case_court
• case_law_suspends_application_of_resource_legal
• date_creation_legacy
• datetime_negotiation
• opinion_advocate_general_joined_to_case_court
• resource_legal_corrects_resource_legal
• resource_legal_id_obsolete_document
• resource_legal_incorporates_resource_legal
• resource_legal_information_miscellaneous
• resource_legal_number_sequence_celex
• resource_legal_uses_originally_language
• summary_case_law_id_celex
• summary_summarizes_work
• work_cites_work
• work_created_by_agent
• work_date_creation
• work_datetime_transmission
• work_embargo
• work_id_obsolete_notice
• work_is_logical_successor_of_work
• work_is_member_of_complex_work
• work_part_of_collection_document
• work_part_of_dossier
• work_part_of_event
• work_related_to_work
• work_title

For a full field reference (type, cardinality, source, examples), see FIELDS.md bundled with this dataset.

How to cite

If you use this dataset in academic work, please cite both the source software and the underlying CELLAR / InfoCuria publications:

@misc{cjeu-opendata-2026,
  title  = {CJEU / CELLAR Case Law},
  author = {Wicker, David},
  year   = {2026},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/datasets/davidwickerhf/cjeu-opendata}},
}

@software{cellar-extractor,
  title  = {cellar-extractor: a Python toolkit for CJEU corpus extraction},
  author = {{Maastricht Law \& Tech}},
  url    = {https://github.com/maastrichtlawtech/cellar-extractor},
}

License

• Code (extraction + consolidation pipeline): Apache-2.0.
• Dataset content: EU institutional content. Court judgments are public-domain in the EU; metadata and citation graph are derived from the EU's CELLAR open-data programme.

Re-use is unconstrained for research, teaching, and commercial purposes; attribution to the source (CELLAR / curia.europa.eu) is appreciated.