LEGEX Goldensets: Expert-Coded Review-Table Annotations

This repository contains the expert-coded gold annotations for the LEGEX benchmark of civil-judgment review-table extraction. 1,446 judgments across 19 jurisdictions have been annotated by hand against a shared 14-field schema covering monetary outcomes, cost allocation, party structure, and industry classification. Including independent secondary re-annotations, the release holds 1,691 annotation rows.

Dataset summary

Legal review-table systems are increasingly used to extract structured facts from judgments, but there is little public evidence on their reliability in cross-jurisdictional legal settings. LEGEX is an expert-coded benchmark for civil-judgment review-table extraction. The current release contains an eight-jurisdiction core benchmark with at least 100 judgments per jurisdiction, each coded by two or more independent experts on an 18–30 case overlap and eleven preview jurisdictions for testing schema portability. The re-annotated overlap enables inter-annotator agreement (IAA) analysis on the core set.

Schema

Each line in data/<cc>/goldenset_<cc>.jsonl is a JSON object with these keys:

Key	Type	Description
case_id	string	Identifier within the source database.
link	string	URL to the original judgment.
full_text	string | null	Full judgment text used as model input.
legal_subject_judgement	string	Short English subject of the case. Acts as the "this row has been substantively reviewed" marker.
trial_start_date	YYYY-MM-DD | null	Trial start date.
trial_end_date	YYYY-MM-DD | null	Decision date.
dispute_value_nominal	string | null	Amount in dispute as a string (e.g. "150000") or the literal "nonpecuniary".
Currency_dispute_value_nominal	string | null	ISO-4217 currency code.
plaintiff_loosing_share	number [0, 1] | null	Plaintiff's losing share.
court_cost_awarded_nominal	number | null	Court fees awarded.
Currency_court_cost_awarded_nominal	string | null	ISO-4217 currency code.
party_compensation_awarded_nominal	number | null	Party compensation awarded.
Currency_party_compensation_awarded_nominal	string | null	ISO-4217 currency code.
plaintiffs_all_count	integer | null	Number of plaintiffs.
defendants_all_count	integer | null	Number of defendants.
plaintiff_no1_ISIC1_industry_category	string | null	ISIC Section A–U for the first plaintiff.
defendant_no1_ISIC1_industry_category	string | null	ISIC Section A–U for the first defendant.

A row is included when legal_subject_judgement is populated (the marker the annotator used to flag a row as substantively reviewed). Cells left empty by the annotator are stored as null. In the raw JSONL the trial dates are YYYY-MM-DD strings. The declared dataset features type them as date32, so the datasets library returns them as date objects.

Traceability fields

Every record additionally carries provenance fields so each value is auditable:

Key	Type	Description
annotator_id	string	Pseudonymous, salted-hash id of the coder, stable per person across jurisdictions.
comment	string | null	Human-readable note listing every field changed during sanitization, e.g. "The trial_end_date was sanitized from '2026_03_19' to '2026-03-19'.", null when the record needed no change.
original_input	string	JSON object mapping each changed field to its 1:1 pre-sanitization value, "{}" when nothing changed. This lets any consumer reconstruct the raw value.

Primary and secondary annotations

Jurisdictions that received an independent re-annotation contain both the primary annotator's rows and the secondary annotators' rows, distinguished by annotator_id. Within each file all primary rows come first and re-annotation rows are appended after them, so the first row per case_id is the primary gold annotation. For benchmark evaluation, deduplicate to the primary rows. For inter-annotator agreement, group the duplicated case_ids by annotator_id.

Jurisdictions

Highest civil court per jurisdiction

Jurisdiction	Highest court (civil jurisdiction)	Language(s)
Armenia	Court of Cassation (Civil Chamber)	Armenian
Australia	High Court of Australia	English
Belgium	Cour de cassation / Hof van Cassatie	French, Dutch
Brazil	Superior Tribunal de Justiça	Portuguese
France	Cour de cassation	French
Georgia	Supreme Court of Georgia	Georgian
Germany	Bundesgerichtshof	German
Hong Kong	Court of Final Appeal	English, Chinese
India	Supreme Court of India	English
Nepal	Supreme Court of Nepal	Nepali
New Zealand	Court of Appeal / High Court	English
Philippines	Supreme Court of the Philippines	English, Filipino
Serbia	Supreme Court of Cassation	Serbian
Singapore	Supreme Court (incl. SICC)	English
Spain	Tribunal Supremo (Sala de lo Civil)	Spanish
Switzerland	Federal Supreme Court	German, French, Italian
Taiwan	Supreme Court	Chinese
United Kingdom	Supreme Court of the United Kingdom	English
United States	Supreme Court of the United States	English

Data sources, judgment counts and annotation rows

# judgments counts unique expert-coded cases, # rows additionally counts independent secondary re-annotations of the same cases.

Jurisdiction	Data source	Time span	# judgments	# rows	Annotators
Armenia	cassationcourt.am	2024 – 2026	58	58	1
Australia	HF isaacus/high-court-of-australia-cases	2015 – 2025	30	30	1
Belgium	juportal.be	2015 – 2025	30	30	1
Brazil	scon.stj.jus.br	2002 – 2026	130	148	2
France	Judilibre (PISTE) API	2015 – 2025	25	25	1
Georgia	supremecourt.ge	2026 only	112	161	3
Germany	HF openlegaldata/court-decisions-germany	2015 – 2022	130	150	2
Hong Kong	legalref.judiciary.hk	2015 – 2026	10	10	1
India	AWS Open Data Indian Supreme Court	2017 – 2023	24	24	1
Nepal	nkp.gov.np	2021 – 2025	130	130	1
New Zealand	justice.govt.nz JDO API	2015 – 2025	29	29	1
Philippines	elibrary.judiciary.gov.ph	2024 – 2025	10	10	1
Serbia	vrh.sud.rs	2023 – 2025	26	26	1
Singapore	sgcaselaw.com	2025 – 2026	124	172	3
Spain	poderjudicial.es	2026 only	58	58	1
Switzerland	HF voilaj/swiss-caselaw	2024 – 2025	130	160	2
Taiwan	judgment.judicial.gov.tw	2026 only	130	150	2
United Kingdom	caselaw.nationalarchives.gov.uk	2015 – 2025	130	160	2
United States	HF free-law/Caselaw_Access_Project	2020 – 2026	130	160	2

Total: 1,446 expert-coded judgments, 1,691 annotation rows.

Core benchmark (≥ 100 expert-coded judgments, coded by ≥ 2 independent annotators): Brazil, Georgia, Germany, Singapore, Switzerland, Taiwan, United Kingdom, United States. In each core jurisdiction 18–30 cases were independently re-annotated by one or two additional experts. Preview jurisdictions: the remaining eleven (Armenia, Australia, Belgium, France, Hong Kong, India, Nepal, New Zealand, Philippines, Serbia, Spain). Nepal reaches 130 cases but is single-annotated and therefore remains preview.

Loading

from datasets import load_dataset

# Single jurisdiction
ds = load_dataset("legexbenchmark/goldensets", split="switzerland")
ds = load_dataset("legexbenchmark/goldensets", split="united_states")

# All jurisdictions
from datasets import concatenate_datasets
splits = [
    "armenia", "australia", "belgium", "brazil", "france", "georgia",
    "germany", "hong_kong", "india", "nepal", "new_zealand", "philippines",
    "serbia", "singapore", "spain", "switzerland", "taiwan",
    "united_kingdom", "united_states",
]
all_rows = concatenate_datasets([
    load_dataset("legexbenchmark/goldensets", split=s) for s in splits
])

Limitations

• Sample sizes: Ten of the nineteen jurisdictions have fewer than 100 expert-coded judgments, and Nepal is single-annotated. These eleven are marked as preview and intended for schema-portability checks rather than per-jurisdiction performance claims.
• Time coverage: Source databases vary considerably (Georgia, Spain and Taiwan are 2026-only because earlier years were not freely available).
• Schema portability: the 14 fields were designed against civil judgments in common-law and Western European civil-law systems. Some fields (e.g. plaintiff_loosing_share, ISIC categorisation) may not be a natural fit for every jurisdiction.
• Court selection: For Brazil, there are judgements that are not from the highest possible court. This can be inferred with the case id.
• Full-text quality: Text was extracted from heterogeneous sources (HTML, PDF, API JSON). Layout artefacts and OCR errors are possible.

Citation

Anonymous submission to the ICML 2026 AI for Law workshop. Citation block will be added after the camera-ready release.

License

MIT.