File size: 3,514 Bytes
287e418
 
6f5156a
 
 
 
 
 
 
287e418
6f5156a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
---
license: mit
pretty_name: "LEGEX Code: Scrapers, Inference and Evaluation Pipeline"
tags:
  - legal
  - benchmark
  - code
  - llm-evaluation
  - information-extraction
---

# LEGEX Code, Scrapers, Inference and Evaluation Pipeline

Python source for the LEGEX benchmark of civil-judgment review-table
extraction. This repository contains:

- Scrapers for 19 jurisdictions (per-court HTML / API / HuggingFace
  pull) in [`legex/scrapers/`](legex/scrapers/).
- Inference pipeline that calls Harvey, Gemini and OpenAI APIs against
  a schema-constrained 14-field review table
  ([`legex/inference.py`](legex/inference.py),
  [`legex/harvey.py`](legex/harvey.py),
  [`legex/models/classification.py`](legex/models/classification.py)).
- Evaluation that compares system outputs against expert-coded gold
  cells ([`legex/evaluation.py`](legex/evaluation.py)), aggregates across
  jurisdictions ([`legex/analysis.py`](legex/analysis.py)), and renders
  paper tables ([`legex/quant_results.py`](legex/quant_results.py)).
- Conversion script [`convert_goldenset_to_jsonl.py`](convert_goldenset_to_jsonl.py)
  — turns the source XLSX goldensets into the JSONL format used by
  [`legexbenchmark/goldensets`](https://huggingface.co/datasets/legexbenchmark/goldensets).


## Setup

```bash
git clone https://huggingface.co/datasets/legexbenchmark/code legex-code
cd legex-code
uv sync
cp .env.template .env
```

Required tokens depend on which scrapers / models you run, see
[`.env.template`](.env.template).

## End-to-end workflow

```bash
# Acquire raw judgments per jurisdiction.
uv run legex-run

# Run inference for one system on one jurisdiction, Harvey has do be done separately as this is a commercial tool
uv run legex-classify --country us --model gpt-5.4-mini --full_text

# Evaluate one system on one jurisdiction.
uv run legex-evaluate --country us --system gpt

# Aggregate across all 12 evaluated jurisdictions and 3 systems.
uv run legex-analysis --out data/analysis

# Render the paper-headline LaTeX table.
uv run legex-quant-results \
    --input data/analysis/per_country_per_column.csv \
    --out   data/analysis/quant_results.tex
```

To evaluate against the published goldensets and inference outputs, pull
the two data repos into the expected layout:

```bash
huggingface-cli download legexbenchmark/goldensets         --repo-type dataset --local-dir data --include "data/*"
huggingface-cli download legexbenchmark/inference-results  --repo-type dataset --local-dir data --include "data/*"
# After these, data/<cc>/ contains goldenset_<cc>.jsonl + inference_*.csv
uv run legex-analysis --out data/analysis
```

## CLI entrypoints

| Command | Module | Purpose |
|---|---|---|
| `legex-run` | `legex.main:main` | Top-level scrape + filter + sample pipeline. |
| `legex-classify` | `legex.inference:main` | Run an LLM over the goldenset and write predictions to CSV. |
| `legex-harvey-ingest` | `legex.harvey:main` | Ingest a Harvey Vault Review export into the per-jurisdiction CSV format. |
| `legex-evaluate` | `legex.evaluation:main` | Per-country, per-field bucket counts and recall / hallucination. |
| `legex-analysis` | `legex.analysis:main` | Cross-jurisdiction analysis → CSV + LaTeX tables. |
| `legex-quant-results` | `legex.quant_results:main` | Paper-headline summary from the analysis CSV. |
| `legex-pdf` | `legex.pdf_export.cli:main` | Render per-row PDFs from a goldenset workbook. |
| `legex-plots` | `legex.plots:main` | Plot helpers used in the paper. |

## License

MIT.