language:
- en
- zh
license: mit
pretty_name: LegalScope
task_categories:
- question-answering
- text-generation
tags:
- legal
- legal-reasoning
- benchmark
- llm-evaluation
- exam-to-case-transfer
- chinese-civil-judgments
- text
size_categories:
- n<1K
LegalScope
LegalScope is a public metadata preview for a benchmark studying whether strong public legal-exam performance transfers to real-case legal analysis.
The benchmark pairs scalable public legal-exam tasks with lawyer-reviewed, de-identified Chinese civil-judgment analysis. This Hugging Face dataset page is kept intentionally small and public-safe: it mirrors the public metadata from the GitHub repository and does not publish the full benchmark workbook, full prompts, reference answers, model outputs, human review sheets, paper draft, or non-de-identified legal source materials.
GitHub repository: https://github.com/EternWang/LegalScope
Benchmark Snapshot
| Component | Count |
|---|---|
| Public legal-exam questions | 868 |
| Real-case issue-stance prompts | 76 |
| Total dataset items | 944 |
| De-identified Chinese civil judgments | 15 |
| Legal issues extracted from judgments | 38 |
| Model groups evaluated | 20 |
| Public-exam model responses | 17,360 |
| Real-case model responses | 1,520 |
| Total dataset model responses | 18,880 |
| Human-validation responses | 1,800 |
Files
data/metadata/dataset_summary.json
data/metadata/model_groups.csv
data/metadata/source_composition.csv
dataset_summary.json contains the public release boundary, benchmark module
counts, aggregate response counts, and transfer-summary metrics.
model_groups.csv lists the 20 evaluated model groups.
source_composition.csv summarizes public-exam jurisdiction/domain composition
and real-case legal-domain composition.
Evaluation Design
LegalScope has two tracks:
- Public legal exams: open-ended public legal-exam questions scored with a reference-aware 0-4 answer-match protocol.
- Chinese real cases: issue-stance prompts derived from de-identified Chinese civil judgments, scored across citation relevance, constraint extraction, and argument validity.
Human validation covers 80 public-exam items and 10 real-case prompts across the same 20 model groups, for 1,800 human-validation responses.
Intended Use
This public preview is intended for:
- understanding the benchmark design and release boundary;
- inspecting public-safe metadata and source composition;
- citing the public LegalScope artifact;
- coordinating access to the private benchmark materials where appropriate.
It is not intended for legal advice, legal decision support, training legal decision systems, or reconstructing withheld private data.
Public Release Boundary
This dataset repository does not include:
- the full benchmark workbook;
- complete prompt matrices;
- reference answers;
- model-output matrices;
- human review sheets or adjudication notes;
- paper draft, review PDF, or LaTeX source;
- non-de-identified judgments or private source documents.
The public files are sufficient to document scope, counts, metadata, and release boundaries, but they are not sufficient to reconstruct the full benchmark.
Citation
If you use this public preview, cite the GitHub repository and the released paper once a stable paper citation is available.
@misc{wang2026legalscope,
title = {LegalScope: Measuring Exam-to-Case Transfer in LLM Legal Reasoning},
author = {Wang, Hongyu},
year = {2026},
howpublished = {\url{https://github.com/EternWang/LegalScope}},
note = {Public metadata preview}
}
Disclaimer
LegalScope is a research benchmark for model evaluation. It is not legal advice, a legal research product, or a substitute for jurisdiction-specific legal review.