metadata

configs:
  - config_name: bge_level
    data_files:
      - split: train
        path: bge_level/train.parquet
      - split: validation
        path: bge_level/val.parquet
      - split: test
        path: bge_level/test.parquet
  - config_name: regeste_level
    data_files:
      - split: train
        path: regeste_level/train.parquet
      - split: validation
        path: regeste_level/val.parquet
      - split: test
        path: regeste_level/test.parquet
  - config_name: text_level
    data_files:
      - split: train
        path: text_level/train.parquet
      - split: validation
        path: text_level/val.parquet
      - split: test
        path: text_level/test.parquet
task_categories:
  - translation
language:
  - de
  - fr
  - it
tags:
  - legal
pretty_name: Swiss Decision Summary Translations
size_categories:
  - 10K<n<100K

Dataset Card for Swiss Decision Summary Translations

This dataset card aims to describe a dataset from summaries of Swiss leading court decisions (also known as "BGE" or "Bundesgerichtsentscheide") from the Swiss Supreme Court: https://bger.ch. It is a parallel dataset that offers translations at three different levels: the entire summary (bge_level), the regeste level (regeste_level), and the text level (text_level).

Dataset Details

Dataset Description

This dataset consists of summaries of leading Swiss Supreme Court decisions. These summaries are provided in three languages (German, French, Italian) and are available at three distinct levels of granularity:

bge_level: The entire summary of the court decision.
regeste_level: A summary focused on the core legal issue.
text_level: A further detailed extraction of specific legal statements.

The dataset can be used for various natural language processing tasks such as multilingual text alignment, machine translation, and legal text analysis.

Curated by: Joel Niklaus
Funded by [optional]: Swiss Federal Supreme Court
Shared by [optional]: Joel Niklaus
Language(s) (NLP): German (de), French (fr), Italian (it)
License: [More Information Needed]

Dataset Sources [optional]

Repository: https://github.com/JoelNiklaus/SwissLegalTranslations
Paper [optional]: SwiLTra-Bench: The Swiss Legal Translation Benchmark
Demo [optional]: [More Information Needed]

Uses

Direct Use

This dataset can be used for:

Multilingual text alignment and translation tasks between German, French, and Italian.
Legal text analysis and summarization, with a focus on Swiss law.
Language model training or fine-tuning on legal domain-specific text.

Out-of-Scope Use

The dataset is not suitable for tasks unrelated to legal text processing or without consideration for the specific legal context of Swiss Supreme Court decisions.

Dataset Structure

The dataset is structured into three levels:

bge_level: Contains full summaries of court decisions in German, French, and Italian.
- Fields:
  - bge: Case identifier.
  - year: Year of the decision.
  - volume: Volume number of the decision.
  - pageNumber: Page number of the decision.
  - de_bgeText: Full summary in German.
  - fr_bgeText: Full summary in French.
  - it_bgeText: Full summary in Italian.
regeste_level: Contains regeste summaries (core legal issues) in the three languages.
- Fields:
  - bge: Case identifier.
  - year: Year of the decision.
  - volume: Volume number of the decision.
  - pageNumber: Page number of the decision.
  - regesteNumber: Number assigned to the regeste.
  - de_regesteTitle: Regeste title in German.
  - fr_regesteTitle: Regeste title in French.
  - it_regesteTitle: Regeste title in Italian.
  - de_regesteText: Regeste text in German.
  - fr_regesteText: Regeste text in French.
  - it_regesteText: Regeste text in Italian.
text_level: Contains more detailed text extracts from the decisions, available in all three languages.
- Fields:
  - bge: Case identifier.
  - year: Year of the decision.
  - volume: Volume number of the decision.
  - pageNumber: Page number of the decision.
  - regesteNumber: Number assigned to the regeste.
  - textNumber: Number assigned to the specific text extract.
  - de_text: Text extract in German.
  - fr_text: Text extract in French.
  - it_text: Text extract in Italian.

Example Code

from datasets import load_dataset

# Load the datasets for each level
bge_dataset = load_dataset('joelniklaus/SwissDecisionSummaryTranslations', name='bge_level', trust_remote_code=True)
regeste_dataset = load_dataset('joelniklaus/SwissDecisionSummaryTranslations', name='regeste_level', trust_remote_code=True)
text_dataset = load_dataset('joelniklaus/SwissDecisionSummaryTranslations', name='text_level', trust_remote_code=True)

# Print datasets for inspection
print(bge_dataset, regeste_dataset, text_dataset)

Dataset Creation

Curation Rationale

The dataset was created to enable the development of multilingual models for the legal domain, specifically focused on Swiss law. By providing translations at multiple levels of granularity, it supports a range of use cases from high-level summaries to detailed legal text analysis.

Source Data

The dataset is derived from official Swiss Supreme Court decision summaries, which are publicly available on bger.ch.

Data Collection and Processing

The summaries were extracted and processed from official court records, normalized across the three languages (German, French, Italian), and structured into the different levels (bge_level, regeste_level, text_level). The dataset was cleaned and formatted into JSONL format for ease of use.

Who are the source data producers?

The source data is produced by the Swiss Federal Supreme Court, which publishes leading court decisions in multiple languages (German, French, and Italian).

Annotations [optional]

No additional annotations have been added beyond the original translations provided by the Swiss Supreme Court.

Personal and Sensitive Information

The dataset does not contain any personal, sensitive, or private information as it consists solely of legal summaries from public court decisions.

Bias, Risks, and Limitations

The dataset may reflect biases inherent in the Swiss legal system and judicial processes. Users should be aware that the dataset only represents leading decisions from the Swiss Supreme Court, which may not reflect the entirety of Swiss law or legal perspectives.

Recommendations

Users should consider the legal context and jurisdiction-specific nature of the dataset when applying models trained on this data. The dataset is best used for tasks related to legal language and multilingual processing in the legal domain.

Citation [optional]

BibTeX:

If you find this dataset helpful, feel free to cite our publication SwiLTra-Bench: The Swiss Legal Translation Benchmark:

@misc{niklaus2025swiltrabenchswisslegaltranslation,
      title={SwiLTra-Bench: The Swiss Legal Translation Benchmark}, 
      author={Joel Niklaus and Jakob Merane and Luka Nenadic and Sina Ahmadi and Yingqiang Gao and Cyrill A. H. Chevalley and Claude Humbel and Christophe Gösken and Lorenzo Tanzi and Thomas Lüthi and Stefan Palombo and Spencer Poff and Boling Yang and Nan Wu and Matthew Guillod and Robin Mamié and Daniel Brunner and Julio Pereyra and Niko Grupen},
      year={2025},
      eprint={2503.01372},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2503.01372}, 
}

APA:

Niklaus, J., Merane, J., Nenadic, L., Ahmadi, S., Gao, Y., Chevalley, C. A. H., Humbel, C., Gösken, C., Tanzi, L., Lüthi, T., Palombo, S., Poff, S., Yang, B., Wu, N., Guillod, M., Mamié, R., Brunner, D., Pereyra, J., & Grupen, N. (2025). SwiLTra-Bench: The Swiss Legal Translation Benchmark. arXiv. https://arxiv.org/abs/2503.01372

Glossary [optional]

BGE: Bundesgerichtsentscheide (leading court decisions from the Swiss Supreme Court).
Regeste: A concise summary of the legal issue(s) in a court decision.

More Information [optional]

[More Information Needed]

Dataset Card Authors

Joel Niklaus

Dataset Card Contact

Joel Niklaus