This is an enhanced version of the Japanese Bar Examination QA dataset, containing question-answer pairs from the Japanese Bar Examination (司法試験, Shihou Shiken) spanning from 2015 to 2024. The dataset covers three major areas of Japanese law: Criminal Law (刑法), Constitutional Law (憲法), and Civil Law (民法).

This version (v2) includes significant improvements in data extraction and processing, with enhanced support for complex question formats and better contextual information.

What's New in v2

Key Improvements over v1

Enhanced Question Extraction: Improved XML parsing that correctly handles complex question formats and edge cases
Lead-in Context Support: Added support for questions with contextual lead-in text (見解, opinions) using the new lead_in field
Remark Annotations: Separated important annotations from instructions into a dedicated remark field for better clarity
Improved Instruction Clarity: More explicit and standardized instruction text for better understanding
Increased Dataset Size: 3,464 questions (vs. 2,846 in v1) - a 21.7% increase
Better Data Quality: Fixed mapping issues between questions and answers, eliminated gold label mismatches

Technical Enhancements

Lead-in Text Extraction: 211 questions (6.1%) now include contextual lead-in text that was previously lost
Remark Annotations: 53 questions (1.5%) include important annotations separated from instructions
Comprehensive Instructions: 3464 questions (100.0%) have clear, standardized instructions
Extended Coverage: Additional questions from years that were previously incompletely processed

Dataset Details

Total Questions: 3,464
Time Period: 2015-2024 (10 years)
Languages: Japanese
Task: Binary classification (True/False questions)
Domains: Legal, specifically Japanese law

Data Fields

id: Unique identifier for each question (format: year-subject-question_number-sub_number)
year: Examination year in Japanese era format (e.g., "令和元年", "平成２７年")
subject: Legal subject in English (Criminal Law, Constitutional Law, Civil Law)
subject_jp: Legal subject in Japanese (刑法, 憲法, 民法)
theme: Specific legal topic/theme in Japanese
question_type: Type of question (CSQ, MRQ, TFQ)
instruction: Clear instructions for answering the question
lead_in: [NEW] Contextual lead-in text providing background information (e.g., legal opinions, case scenarios)
remark: [NEW] Important annotations and clarifications separated from instructions
question: Question text in Japanese
label: Original label (Y for True, N for False)
answer: Standardized answer (True/False)

Data Splits

The dataset is split into three parts:

Train: 2771 questions (80%)
Validation: 346 questions (10%)
Test: 347 questions (10%)

Subject Distribution

Subject	Count	Percentage
Civil Law	1,998	57.7%
Criminal Law	811	23.4%
Constitutional Law	655	18.9%

Answer Distribution

Answer	Count	Percentage
False	1,814	52.4%
True	1,650	47.6%

Question Types

Type	Description	Count	Percentage
CSQ	Choice of Single Question	2,396	69.2%
MRQ	Multiple Response Question	593	17.1%
TFQ	True/False Question	475	13.7%

New Features in v2

Lead-in Context

211 questions (6.1%) now include lead-in text that provides essential context:

Legal Opinions (見解): Different theoretical perspectives on legal issues
Case Scenarios (事例): Factual backgrounds for legal analysis
Comparative Views: Multiple viewpoints on legal interpretations

Example with lead-in:

{
  "id": "令和６年-刑法-11-1",
  "lead_in": "【見解】A説:行為者が、やろうと思えばできたが中止した場合を中止犯とし...",
  "question": "A説の立場からは、中止行為が反省・悔悟等の自己の行為に対する否定的な感情に基づく場合に限り、中止犯が成立する。",
  "instruction": "次の各【見解】に対して記述が正しいか"
}

Remark Annotations

53 questions (1.5%) include important clarifications:

Assumptions: Conditions to assume when answering (e.g., "Bは、無資力であり...")
Legal Frameworks: Specific legal provisions to consider
Contextual Notes: Additional information for proper interpretation

Example with remark:

{
  "id": "令和５年-民法-18-1",
  "instruction": "判例の趣旨に照らし記述が正しいか",
  "remark": "Bは、無資力であり、各行為が債権者を害することを知っていたものとする。"
}

Usage Examples

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("nguyenthanhasia/japanese-bar-exam-qa-v2")

# Access different splits
train_data = dataset["train"]
validation_data = dataset["validation"]
test_data = dataset["test"]

Example Data Point

{
    "id": "令和３年-民法-24-3",
    "year": "令和３年",
    "subject": "Civil Law",
    "subject_jp": "民法",
    "theme": "贈与",
    "question_type": "CSQ",
    "instruction": "判例の趣旨に照らし記述が正しいか",
    "lead_in": "",
    "remark": "",
    "question": "受贈者は,贈与契約が書面によらない場合であっても,履行の終わっていない部分について贈与契約を解除することができない。",
    "label": "N",
    "answer": "False"
}

Applications

This enhanced dataset can be used for:

Advanced Legal NLP Research: With improved context through lead-in text and remarks
Legal Reasoning Systems: Better understanding of legal argumentation patterns
Educational AI Tools: More comprehensive study materials for law students
Legal AI Development: Training models with richer contextual information
Cross-lingual Legal Studies: Enhanced comparative legal research capabilities

Improvements from v1

Data Quality Enhancements

Fixed Gold Label Issues: Resolved answer mapping problems identified in v1
Enhanced Extraction: Improved handling of complex question formats
Better Parsing: More robust XML processing with error handling
Standardized Instructions: Clearer and more consistent instruction text

Coverage Expansion

Additional Questions: 618 new questions extracted from previously incomplete processing
Better Representation: More balanced coverage across years and subjects
Contextual Information: Previously lost contextual information now preserved

Data Source

The data originates from official Japanese Bar Examination questions administered by the Ministry of Justice of Japan. This version includes enhanced processing of the original examination materials with improved extraction techniques.

Version History

v2.0: Enhanced extraction with lead-in context, remark annotations, and improved data quality
v1.0: Initial release with basic question-answer pairs

Considerations for Use

Limitations

The dataset is in Japanese and requires Japanese language processing capabilities
Questions are based on Japanese legal system and may not be applicable to other legal systems
Some questions may require deep understanding of Japanese legal precedents and context
The dataset reflects the legal understanding and interpretations at the time of each examination

Ethical Considerations

This dataset is intended for research and educational purposes
Users should be aware that legal interpretations may change over time
The dataset should not be used as a substitute for professional legal advice
Proper attribution should be given to the original source (Japanese Ministry of Justice)

Citation

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

@misc{japanese_bar_exam_qa_v2_2025,
  title        = {Japanese Bar Examination QA Dataset v2},
  author       = {Fumihito Nishino and Nguyen Ha Thanh and Ken Satoh},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/datasets/nguyenthanhasia/japanese-bar-exam-qa-v2}},
  note         = {Publisher: Hugging Face}
}

License

This dataset is released under the CC BY 4.0 license.

The underlying content originates from official government sources that are in the public domain. Please refer to the original source for their terms of use.

Contact

For questions or issues regarding this dataset, please open an issue in the dataset repository.

Related Datasets

japanese-bar-exam-qa v1: Original version of this dataset

Downloads last month: 208