Hugging Face's logo

Datasets:
legalontech
/
Legal-Rikai-Open-Benchmark

Tasks:
Text Generation
Modalities:
Text
Formats:
json
Languages:
Japanese
Size:
< 1K
ArXiv:
Tags:
legal
benchmark
japanese
contract-revision
Libraries:
Datasets
pandas
License:
Dataset card Data Studio Files
xet
Community
1

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

To request access to this dataset, please provide the following information and agree to the Terms of Use. The submitted information may be used to contact you regarding your usage of the dataset and compliance with the Terms.

Log in or Sign Up to review the conditions and access this dataset content.

Dataset Summary

Legal Rikai: Open Benchmark is a dataset designed to evaluate LLMs on complex Japanese corporate legal tasks. Unlike existing benchmarks that focus on short-answer QA or classification, this benchmark targets long-form, structured outputs and document-level editing capabilities essential for practical legal work. The dataset consists of four distinct tasks that emulate real-world workflows handled by legal departments, such as explaining legal amendments and revising contracts based on statutes, specific requirements, or risk mitigation needs. All data instances were created by legal professionals under the supervision of qualified lawyers.

Notice on Data Formatting

The dataset released here represents the finalized version. Due to formatting issues present in the data used during the experiments reported in the paper, some details may differ between the paper’s experimental data and this official release.

Task Description

The benchmark comprises four sub-tasks, each containing 25 samples.

1. Legal Amendment Explanation (AmendExp)

  • Goal: Compare pre- and post-amendment statutes to generate a summary of changes and explain their impact on contracts.
  • Input: Pre-amendment statute, Post-amendment statute.
  • Output: Summary of amendments, Impact on contracts.

2. Statute-driven Contract Revision (StatRev)

  • Goal: Revise an existing contract to comply with a newly amended statute.
  • Input: Pre-amendment statute, Post-amendment statute, Pre-amendment compliant contract.
  • Output: A revised contract compliant with the current statute.

3. Requirement-driven Contract Revision (ReqRev)

  • Goal: Revise a contract based on specific instructions from stakeholders (e.g., counterparty requests).
  • Input: Contract template, Revision instructions.
  • Output: Revised contract.

4. Risk-driven Contract Revision (RiskRev)

  • Goal: Revise a contract to mitigate abstract legal risks (e.g., "revise to be more favorable to the contractor").
  • Input: Contract template, Risk mitigation instructions.
  • Output: Revised contract with minimized legal risks.

Dataset Structure

DataRecords

AmendExp

  • id (int32): Unique sample ID (0-24).
  • reference (string): The ground truth text
  • source_statute_before_amendment (string): Full text of the statute before revision.
  • source_statute_after_amendment (string): Full text of the statute after revision.
  • meta (dict):
    • source_statute_before_amendment_url: Source URL for the old statute.
    • source_statute_after_amendment_url: Source URL for the new statute.

StatRev

  • id (int32): Unique sample ID (0-24).
  • source (string): The original input contract (Markdown).
  • reference (string): The ground truth revised contract (Markdown).
  • source_statute_before_amendment (string): Full text of the statute before revision.
  • source_statute_after_amendment (string): Full text of the statute after revision.
  • meta (dict):
    • source_statute_before_amendment_url: Source URL for the old statute.
    • source_statute_after_amendment_url: Source URL for the new statute.

Note: In some cases, the correct response may be to leave the contract unchanged (i.e., source and reference may be identical). This represents cases where the instruction does not require any modification to the original contract.

ReqRev & RiskRev: These tasks involve contract editing based on instructions.

  • id (int32): Unique sample ID (0-24).
  • instruction (string): The revision instructions specifying how to modify the contract.
  • source (string): The original input contract (Markdown).
  • reference (string): The ground truth revised contract (Markdown).

Model Outputs

Contains the raw text generated by the models.

  • id (int32): Unique sample ID.
  • gpt-5 (string): Output generated by GPT-5.
  • claude-opus-4-1 (string): Output generated by Claude Opus 4.1.
  • gemini-2.5-pro (string): Output generated by Gemini 2.5 Pro.

Note: For the Statute-driven Contract Revision task, models output only the clauses that need to be modified, not the entire contract.

Usage Example 

from datasets import load_dataset

# Load the input data.
amend_exp_data = load_dataset("legalontech/Legal-Rikai-Open-Benchmark", "amend_exp_data")
stat_rev_data = load_dataset("legalontech/Legal-Rikai-Open-Benchmark", "stat_rev_data")
req_rev_data = load_dataset("legalontech/Legal-Rikai-Open-Benchmark", "req_rev_data")
risk_rev_data = load_dataset("legalontech/Legal-Rikai-Open-Benchmark", "risk_rev_data")

# Load the model outputs.
amend_exp_model_output = load_dataset("legalontech/Legal-Rikai-Open-Benchmark", "amend_exp_model_output")
stat_rev_model_output = load_dataset("legalontech/Legal-Rikai-Open-Benchmark", "stat_rev_model_output")
req_rev_model_output = load_dataset("legalontech/Legal-Rikai-Open-Benchmark", "req_rev_model_output")
risk_rev_model_output = load_dataset("legalontech/Legal-Rikai-Open-Benchmark", "risk_rev_model_output")

# Access the test split data
for sample in req_rev_data["test"]:
    print(sample["id"], sample["instruction"])

Terms of Use

LegalOn Technologies, Inc. (hereinafter referred to as the "Company") provides "LegalRikai: Open Benchmark" to users (hereinafter referred to as the "User"). The User shall agree to these Terms of Use (hereinafter referred to as the "Terms") and use LegalRikai: Open Benchmark in accordance with the Terms. The Terms are subject to change without prior notice. 株式会社LegalOn Technologies（以下「当社」といいます。）が提供する「LegalRikai: Open Benchmark」を利用される方（以下「利用者」といいます。）は、本規約に同意いただき、本規約に従ってLegalRikai: Open Benchmarkをご利用ください。本規約は、事前の通知なく変更される場合があります。

About "LegalRikai: Open Benchmark" 「LegalRikai: Open Benchmark」とは “LegalRikai: Open Benchmark” is a dataset created by the Company primarily for the purpose of benchmarking (performance evaluation) of language models. LegalRikai: Open Benchmark may be updated, changed, or discontinued without prior notice. 「LegalRikai: Open Benchmark」は、当社が作成した、主に言語モデルのベンチマーク（性能評価）のためのデータセットです。LegalRikai: Open Benchmarkは、事前の通知なく更新、変更又は公開終了が行われる場合があります。

Conditions of Use for LegalRikai: Open Benchmark LegalRikai: Open Benchmarkの利用条件

  1. The use of LegalRikai: Open Benchmark under the Terms shall be limited to AI research purposes, such as benchmarking (performance evaluation) of language models. Use for any other purpose is prohibited. 本規約に基づくLegalRikai: Open Benchmarkの利用は、言語モデルのベンチマーク（性能評価）等のAIの研究を目的とする利用に限ります。それ以外の目的で利用することはできません。
  2. The User shall not use LegalRikai: Open Benchmark for any acts that violate, or may violate, laws and regulations or public order and morals. LegalRikai: Open Benchmarkについて、法令または公序良俗に反する行為及びその恐れのある行為に利用することはできません。
  3. If the Company suspects that the User’s use of LegalRikai: Open Benchmark violates the purpose set forth in Paragraph 1, or if the Company reasonably determines that there is a risk of such violation, the Company may refuse to grant a license, or suspend or terminate any license already granted, for all or part of the use of LegalRikai: Open Benchmark under the Terms. 当社は、利用者による LegalRikai: Open Benchmark の利用が第1項に定める目的に反すると疑われる場合、またはそのおそれがあると当社が合理的に判断した場合、本規約に基づく LegalRikai: Open Benchmark の利用の全部又は一部について、利用を許諾せず、又は既に付与した利用の許諾を停止し、若しくは解除することができます。
  4. The User represents and warrants that the User does not fall under the category of Anti-Social Forces and does not have any socially condemnable relationship with Anti-Social Forces. 利用者は、反社会的勢力に該当せず、かつ反社会的勢力と社会的に非難されるべき関係を有しないことを表明し、保証します。
  5. If the User is found to fall under the category of Anti-Social Forces, or if the Company determines that there is a risk thereof, the Company may suspend the license to use LegalRikai: Open Benchmark under the Terms without prior notice. The Company shall not be liable for any damages incurred by the User as a result of such suspension. 当社は、利用者が反社会的勢力に該当すると判明した場合、又はその恐れがあると当社が判断した場合、事前にお知らせすることなく、本規約に基づく LegalRikai: Open Benchmark の利用許諾を停止する場合があります。また、これにより利用者に損害が生じても、当社は一切の責任を負いません。
  6. The license granted under the Terms is non-exclusive. LegalRikai: Open Benchmark is not for the exclusive use of any specific User. 本規約に基づく利用の許諾は、非独占的な許諾です。LegalRikai: Open Benchmarkは、特定の利用者が独占的に利用できるものではありません。
  7. The User may not resell, republish, redistribute, sublicense, or otherwise allow any third party other than the User to use all or any part of LegalRikai: Open Benchmark. LegalRikai: Open Benchmarkの一部又は全部を、転売、再販売、転載、再配布、サブラインセンス及びその他利用者以外に利用させることはできません。
  8. The Company makes no warranty regarding the functionality, completeness, accuracy, fitness for a particular purpose, or any other aspect of LegalRikai: Open Benchmark. LegalRikai: Open Benchmarkの機能性、完全性、正確性、特定の目的の適合性その他一切の保証はされません。

Rights related to LegalRikai: Open Benchmark LegalRikai: Open Benchmarkに係る権利

  1. Copyrights and other intellectual property rights (hereinafter referred to as "Copyrights, etc.") related to LegalRikai: Open Benchmark belong to the Company. By agreeing to the Terms, the User obtains the right to use the dataset, but Copyrights, etc. are not transferred from the Company to the User. LegalRikai: Open Benchmarkに関連する著作権その他の知的財産権（以下「著作権等」といいます。）は当社に帰属します。利用者は、本規約に同意することにより利用する権利を得るものであり、当社から利用者に著作権等が譲渡されるものではありません。
  2. The User is prohibited from claiming rights to, or filing applications for or registering Copyrights, etc. for LegalRikai: Open Benchmark or the data constituting LegalRikai: Open Benchmark. LegalRikai: Open BenchmarkおよびLegalRikai: Open Benchmarkを構成するデータについて権利を主張することおよび著作権等の出願または登録等を禁止します。

Other Provisions その他

  1. In the event of any use found to be in violation of the Terms, the Company may take legal action. 本規約に違反する取り扱いが判明した場合には、法的な措置を取る場合があります。
  2. The Japanese version of the Terms is the official version, and the English translation is for reference purposes only. In the event of any discrepancy between the Japanese and English versions, the Japanese version shall prevail. 本規約は、日本語が正であり、英語訳は参考となります。日本語と英語に齟齬がある場合、日本語の内容が優先されます。
  3. The Terms shall be governed by the laws of Japan. The Tokyo District Court shall have exclusive agreed jurisdiction for the first instance over any disputes related to the Terms. 本規約は、日本法に準拠し、本規約に関する紛争については、東京地方裁判所を第一審の専属的合意管轄裁判所とします。 以上

Citation

If you use this dataset in your research, please cite the following paper:

@inproceedings{legalrikai2025,
  title={LegalRikai: Open Benchmark - A Benchmark for Complex Japanese Corporate Legal Tasks},
  author={Shogo Fujita and Yuji Naraki and Yiqing Zhu and Shinsuke Mori},
  booktitle={arXiv preprint arXiv:2512.11297},
  year={2025}
}
Downloads last month
21