anonymous commited on
Commit ·
916a451
1
Parent(s): 2e511b5
Add card metadata to the bundle README
Browse files
README.md
CHANGED
|
@@ -1,14 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
# LEGEX — Reproduction Bundle (release/icml2026)
|
| 2 |
|
| 3 |
-
LEGEX is an expert-coded benchmark for civil-judgment review-table
|
| 4 |
extraction: judgments from the highest civil courts of 19 jurisdictions,
|
| 5 |
annotated by legal experts on a shared schema (monetary outcomes, cost
|
| 6 |
allocation, party structure, industry classification), against which we
|
| 7 |
evaluate two commercial review-table systems (Harvey, Legora) and two
|
| 8 |
schema-constrained LLM pipelines (Gemini, ChatGPT). This bundle is the
|
| 9 |
-
versioned code + data payload behind the paper LEGEX: An Eight-Jurisdiction
|
| 10 |
-
Benchmark for Legal Review-Table Extraction (ICML 2026 Workshop on AI for
|
| 11 |
-
Law)
|
| 12 |
double-annotated on a 28–30-case overlap) and eleven preview jurisdictions.
|
| 13 |
|
| 14 |
The release consists of three Hugging Face repositories, pinned to the
|
|
@@ -101,18 +115,17 @@ bash scripts/reproduce_paper.sh
|
|
| 101 |
git status data/analysis
|
| 102 |
```
|
| 103 |
|
| 104 |
-
## Unpublished
|
| 105 |
|
| 106 |
The expert-annotated XLSX workbooks (`data/<cc>/Goldenset_*_final*.xlsx`)
|
| 107 |
-
behind the goldensets
|
| 108 |
metadata, and the re-annotation assignment structure could identify
|
| 109 |
-
individual annotators. The gold labels are published in full as
|
| 110 |
-
`goldensets/data/<cc>/goldenset_<cc>.jsonl`
|
| 111 |
-
annotator rows, and every analysis in this bundle runs from those published
|
| 112 |
-
files
|
| 113 |
|
| 114 |
-
|
| 115 |
-
The AAT (Calderon, Reichart & Dror,
|
| 116 |
ACL 2025, [arXiv:2501.10970](https://arxiv.org/abs/2501.10970)) is run with
|
| 117 |
the authors' original implementation via the adapter
|
| 118 |
`scripts/alt_test_reference.py`, which extracts `alt_test()` from the
|
|
@@ -137,7 +150,7 @@ The scrapers that built the corpus are included (`legex/scrapers/`, one file
|
|
| 137 |
per jurisdiction; see `docs/Jurisdictions.md` for the per-country catalogue).
|
| 138 |
|
| 139 |
```bash
|
| 140 |
-
cp .env.template .env
|
| 141 |
uv run legex-run
|
| 142 |
```
|
| 143 |
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
pretty_name: "LEGEX Reproduction Bundle: Code, Pipeline, and Analysis Outputs"
|
| 4 |
+
language:
|
| 5 |
+
- en
|
| 6 |
+
tags:
|
| 7 |
+
- legal
|
| 8 |
+
- benchmark
|
| 9 |
+
- llm-evaluation
|
| 10 |
+
- information-extraction
|
| 11 |
+
- reproducibility
|
| 12 |
+
viewer: false
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
# LEGEX — Reproduction Bundle (release/icml2026)
|
| 16 |
|
| 17 |
+
**LEGEX** is an expert-coded benchmark for civil-judgment review-table
|
| 18 |
extraction: judgments from the highest civil courts of 19 jurisdictions,
|
| 19 |
annotated by legal experts on a shared schema (monetary outcomes, cost
|
| 20 |
allocation, party structure, industry classification), against which we
|
| 21 |
evaluate two commercial review-table systems (Harvey, Legora) and two
|
| 22 |
schema-constrained LLM pipelines (Gemini, ChatGPT). This bundle is the
|
| 23 |
+
versioned code + data payload behind the paper *LEGEX: An Eight-Jurisdiction
|
| 24 |
+
Benchmark for Legal Review-Table Extraction* (ICML 2026 Workshop on AI for
|
| 25 |
+
Law); it contains an eight-jurisdiction core benchmark (≥100 judgments each,
|
| 26 |
double-annotated on a 28–30-case overlap) and eleven preview jurisdictions.
|
| 27 |
|
| 28 |
The release consists of three Hugging Face repositories, pinned to the
|
|
|
|
| 115 |
git status data/analysis
|
| 116 |
```
|
| 117 |
|
| 118 |
+
## Unpublished raw data — scope
|
| 119 |
|
| 120 |
The expert-annotated XLSX workbooks (`data/<cc>/Goldenset_*_final*.xlsx`)
|
| 121 |
+
behind the goldensets are **not published**: Microsoft Office embeds author
|
| 122 |
metadata, and the re-annotation assignment structure could identify
|
| 123 |
+
individual annotators. The gold *labels* are published in full as
|
| 124 |
+
`goldensets/data/<cc>/goldenset_<cc>.jsonl` (primary + anonymized secondary
|
| 125 |
+
annotator rows), and every analysis in this bundle runs from those published
|
| 126 |
+
files — the workbooks are not needed for reproduction.
|
| 127 |
|
| 128 |
+
**Alternative Annotator Test (AAT).** The AAT (Calderon, Reichart & Dror,
|
|
|
|
| 129 |
ACL 2025, [arXiv:2501.10970](https://arxiv.org/abs/2501.10970)) is run with
|
| 130 |
the authors' original implementation via the adapter
|
| 131 |
`scripts/alt_test_reference.py`, which extracts `alt_test()` from the
|
|
|
|
| 150 |
per jurisdiction; see `docs/Jurisdictions.md` for the per-country catalogue).
|
| 151 |
|
| 152 |
```bash
|
| 153 |
+
cp .env.template .env # then fill in what you need
|
| 154 |
uv run legex-run
|
| 155 |
```
|
| 156 |
|