legexbenchmark/code

LEGEX — Reproduction Bundle (release/icml2026)

LEGEX is an expert-coded benchmark for civil-judgment review-table extraction: judgments from the highest civil courts of 19 jurisdictions, annotated by legal experts on a shared schema (monetary outcomes, cost allocation, party structure, industry classification), against which we evaluate two commercial review-table systems (Harvey, Legora) and two schema-constrained LLM pipelines (Gemini, ChatGPT). This bundle is the versioned code + data payload behind the paper LEGEX: An Eight-Jurisdiction Benchmark for Legal Review-Table Extraction (ICML 2026 Workshop on AI for Law); it contains an eight-jurisdiction core benchmark (≥100 judgments each, double-annotated on a 28–30-case overlap) and eleven preview jurisdictions.

The release consists of three Hugging Face repositories, pinned to the release/icml2026 branch:

Artifact	Contents	URL
goldensets	Expert gold labels, data/<cc>/goldenset_<cc>.jsonl	https://huggingface.co/datasets/legexbenchmark/goldensets/tree/release/icml2026
inference-results	System predictions, data/<cc>/inference_{harvey,harvey_2,gemini,gpt,legora_1,legora_2}.jsonl, plus the as-run per-column prompts of the commercial runs in prompts/	https://huggingface.co/datasets/legexbenchmark/inference-results/tree/release/icml2026
code	This bundle (package, scripts, shipped analysis outputs)	https://huggingface.co/datasets/legexbenchmark/code/tree/release/icml2026

Layout of this bundle

submission/
├── README.md                       this file
├── pyproject.toml · LICENSE · .env.template
├── legex/                          the LEGEX package, pruned to what the release uses:
│   ├── scrapers/                   one scraper per jurisdiction
│   ├── fulltext/                   full-text acquisition for the goldenset workbooks
│   ├── prompts/                    versioned system prompts (paper runs use v3)
│   ├── inference.py                legex-classify (LLM pipelines)
│   ├── harvey.py                   legex-harvey-ingest (review-table export ingest)
│   ├── evaluation/                 tolerant cell comparator + scoring engine
│   └── analysis/                   aggregates, IAA, AAT rendering, report, plots
├── scripts/                        paper-facing scripts (tables, figures, AAT adapter)
├── tests/                          unit tests (uv run pytest -q)
├── data/analysis/                  SHIPPED analysis outputs (see "Recompute" below)
│   ├── per_country_per_column.csv · per_country.csv · per_column.csv
│   ├── per_tradition.csv · per_language.csv
│   ├── quant_results.tex           paper Table tab:overall
│   ├── paper_tables.tex            tab:metrics-by-jurisdiction / tab:metrics-by-field
│   ├── tables/                     diversity.tex, currency_frequencies.tex, ...
│   ├── figures/                    isic_frequencies.pdf, legal_subject_wordcloud.png
│   ├── iaa/                        ANALYSIS.md, pairwise_agreement.csv, kappa_audit.csv,
│   │                               alt_test_pooled.csv, alt_test_reference_*.csv
│   ├── hallucinations/             hallucination_review_ch.csv (hand-classified error case study)
│   └── quality/                    by_country.csv, by_variable.csv
├── goldensets/                     1:1 copy of legexbenchmark/goldensets (dataset card + data)
├── inference-results/              1:1 copy of legexbenchmark/inference-results
├── convert_goldenset_to_jsonl.py   XLSX gold workbooks -> goldenset_<cc>.jsonl
├── build_inference_jsonl.py        cleaned working JSONL -> inference_<model>.jsonl
├── docs/Jurisdictions.md           per-jurisdiction research catalogue
└── HUGGINGFACE_UPLOAD.md           how this bundle maps onto the three HF repos

Reproduce the paper tables and figures

All commands are run from this directory (the bundle root). Requires uv and Python ≥ 3.11.

uv sync

Table tab:overall (headline metrics). Re-render from the shipped per-cell aggregate and check it is byte-identical to the shipped table:

uv run legex-quant-results --input data/analysis/per_country_per_column.csv --out /tmp/quant_results.tex
diff /tmp/quant_results.tex data/analysis/quant_results.tex

Tables tab:metrics-by-jurisdiction and tab:metrics-by-field.

uv run python scripts/paper_tables.py > /tmp/paper_tables.tex
diff /tmp/paper_tables.tex data/analysis/paper_tables.tex

IAA / kappa / Alternative Annotator Test report (ANALYSIS.md, the source of the paper's tab:iaa and AAT numbers). The report is rendered purely from the shipped CSVs:

uv run legex-analysis-report --iaa-dir data/analysis/iaa --analysis-dir data/analysis --out /tmp/ANALYSIS.md
diff /tmp/ANALYSIS.md data/analysis/iaa/ANALYSIS.md

One-shot reproduction. scripts/reproduce_paper.sh regenerates every artifact under data/analysis/ — IAA CSVs, scoring aggregates, all tables, the diversity/frequency figures, the hallucination shares, and ANALYSIS.md — from the published bundles in goldensets/data/ and inference-results/data/. The run is byte-stable, so a clean git status data/analysis afterwards is the verification that the shipped numbers reproduce:

bash scripts/reproduce_paper.sh
git status data/analysis

There are minimal known deviations from the submitted PDF. The gold labels are normalised during export to JSONL (controlled-token case, thousands separators, Excel dates → ISO), and the paper's tables were computed from the raw annotator workbooks before that normalisation, where, for example, a gold cell coded Nonpecuniary counted against a system answering the schema literal nonpecuniary. Re-scoring on the published gold therefore yields slightly higher values in a few places than the PDF: Philippines recall (+7–9 pp for every system), Spain (≈ +1 pp), tab:overall within ±0.2 pp, and the money/ratio agreement rows of tab:iaa (+1–2 pp). The κ values, pair counts, denominators, the hallucination case study, and all remaining numbers match the PDF. The shipped outputs under data/analysis/ are the published-data regeneration. The deviations could not be resolved due to time constraints and are made transparent here.

Unpublished raw data — scope

The expert-annotated XLSX workbooks (data/<cc>/Goldenset_*_final*.xlsx) behind the goldensets are not published: Microsoft Office embeds author metadata, and the re-annotation assignment structure could identify individual annotators. The gold labels are published in full as goldensets/data/<cc>/goldenset_<cc>.jsonl (primary + anonymized secondary annotator rows), and every analysis in this bundle runs from those published files — the workbooks are not needed for reproduction.

Alternative Annotator Test (AAT). The AAT (Calderon, Reichart & Dror, ACL 2025, arXiv:2501.10970) is run with the authors' original implementation via the adapter scripts/alt_test_reference.py, which extracts alt_test() from the upstream notebook at runtime and feeds it LEGEX data from the published bundles:

git clone https://github.com/nitaytech/AltTest /tmp/AltTest
ALTTEST_DIR=/tmp/AltTest bash scripts/reproduce_paper.sh   # or:
uv run python scripts/alt_test_reference.py --alttest /tmp/AltTest \
    --gold-dir goldensets/data --inference-dir inference-results/data \
    [--countries ge,sg,tw] [--epsilon 0.2] [--out data/analysis/iaa] [--per-field]

Its outputs (alt_test_pooled.csv, alt_test_reference_*.csv) are shipped under data/analysis/iaa/ and rendered into ANALYSIS.md by legex-analysis-report.

Scraping

The scrapers that built the corpus are included (legex/scrapers/, one file per jurisdiction; see docs/Jurisdictions.md for the per-country catalogue).

cp .env.template .env   # then fill in what you need
uv run legex-run

legex-run executes the idempotent pipeline (scrape → filter/sample → fill goldenset → dist) for every registered jurisdiction and skips countries whose access is not set up or whose output already exists. Some sources need manual prerequisites (all keys go into .env, see .env.template):

• Hugging Face (au, ch, de, us): a read token as HF_TOKEN; de and us additionally require accepting the gated-dataset terms of openlegaldata/court-decisions-germany and free-law/Caselaw_Access_Project.
• France: register at https://piste.gouv.fr, subscribe to the Judilibre API, set JUDILIBRE_CLIENT_ID / JUDILIBRE_CLIENT_SECRET.
• New Zealand: copy the WAF cookie from a justice.govt.nz session into NZ_WAF_COOKIE.
• India: legex-india-extract pulls selected PDFs from the AWS Open Data TAR archives (see legex/scrapers/in_.py).

Model inference over the scraped full texts (the paper runs use prompt v3):

uv run legex-classify --model <model> --full_text --prompt_version v3

Tests: uv run pytest -q.

Citation

@inproceedings{legex2026,
  title     = {{LEGEX}: An Eight-Jurisdiction Benchmark for Legal Review-Table Extraction},
  author    = {K{\"o}nig, Adrian and others},
  booktitle = {ICML 2026 Workshop on AI for Law (AI4Law)},
  year      = {2026},
  url       = {https://huggingface.co/legexbenchmark}
}

License

MIT. See LICENSE.