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.

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

Dataset Card for tw-judgment-gist

tw-judgment-gist 是一個收錄中華民國司法院精選判決書之要旨集,合計 31 筆。每筆包含判決書編號字串(jid_str)與完整判決書正文(含裁判字號、日期、案由、當事人、判決理由等),適用於判決書要旨萃取模型之訓練、CPT 或作為 tw-judgment-gist-chat 之原始素材來源。

Dataset Details

Dataset Description

司法院於其公開網站會就具有法律見解重要性之判決書製作「判決要旨」,作為學界與實務界引用之參考。相較於一般判決書,精選判決往往具有以下特徵:

  • 法律見解具突破性或釐清既有爭議;
  • 論理結構完整、事實與理由對應清楚;
  • 在後續實務中被頻繁援引。

本資料集收錄這些精選判決之完整文本,保留原始格式(含裁判字號、日期、案由、當事人欄位、論理段落等),適合作為法律 LLM 學習「精華判決」推理結構之 CPT 語料。對應之 chat 格式版本見 tw-judgment-gist-chat

Dataset Sources

Uses

Direct Use

本資料集主要設計用於:

  • 繁體中文法律 LLM 之持續預訓練(CPT),學習精選判決之論理結構;
  • 判決書要旨萃取任務之訓練素材;
  • 作為 tw-judgment-gist-chat 之原始文本來源;
  • 研究判決書精華段落之語言特徵。

Out-of-Scope Use

本資料集不適用於下列用途:

  • 作為法律諮詢之依據,任何模型輸出不構成法律意見;
  • 作為律師考試準備之完整教材(資料量小,僅 31 筆);
  • 作為當事人身份識別或敏感資訊還原之素材;
  • 非中華民國法律體系之訓練資料。

Dataset Structure

{
  "jid_str": "112 年度台上字第 5096 號",
  "text": "裁判字號:最高法院 112 年度台上字第 5096 號刑事判決\n裁判日期:民國 113 年 06 月 26 日\n裁判案由:違反藥事法\n..."
}
欄位 型別 說明
jid_str string 判決書字號(年度、字別、號數)
text string 判決書完整文本
統計項目 數值
總筆數 31
切分 無(使用者自行切分)

Dataset Creation

Curation Rationale

法律 LLM 若僅以一般判決書為訓練語料,容易被大量例行性、格式重複之判決主導訊號,難以學會真正具價值之法律見解。精選判決書則集中於法律爭議之核心,是高品質之法律論理樣本。本資料集之目標即為提供一個小型但精華之判決書集,作為 CPT 之補強語料。

Source Data

Data Collection and Processing

  • 來源:司法院公開之精選判決書;
  • 處理:保留原始文本結構(字號、日期、案由、當事人、論理段落),以 JSONL 儲存;
  • 格式:每筆僅兩個欄位(jid_strtext)。

Who are the source data producers?

原始判決書由中華民國各級法院之承審法官撰寫,經司法院依其標準選入精選判決。

Annotations

Annotation process

本資料集不包含額外人工標註。

Who are the annotators?

不適用。

Personal and Sensitive Information

判決書為司法院依法公開之文書,當事人姓名之揭露或遮蔽依司法院規則辦理,本資料集未額外匿名化。使用者不得以本資料集進行當事人識別。

Bias, Risks, and Limitations

  • 資料量極小:僅 31 筆,不足以單獨支撐訓練;
  • 領域分布不均:精選判決之領域由司法院選定,可能偏向特定法律領域;
  • 時效性:法律見解會隨法規修訂而變動,歷史精選判決之部分結論可能已過時;
  • 格式雜訊:原始文本含各判決書之編排格式,直接輸入模型時需額外切分或清理。

Recommendations

建議使用者:

  • 將本資料集作為 CPT 之補強語料,與 tw-processed-judgments-14B 等大規模判決書語料共同訓練;
  • 若要做判決書要旨萃取任務,建議使用 tw-judgment-gist-chat 之 QA 格式版本;
  • 關注判決書之時效性,必要時與最新法規交叉比對。

Citation

@misc{tw-judgment-gist,
  title        = {tw-judgment-gist: Curated ROC Judicial Judgments for Legal LLM Training},
  author       = {Liang Hsun Huang},
  year         = {2024},
  howpublished = {\url{https://huggingface.co/datasets/lianghsun/tw-judgment-gist}},
  note         = {31 curated ROC judgments selected by the Judicial Yuan as representative legal opinions.}
}

Dataset Card Authors

Liang Hsun Huang

Dataset Card Contact

Liang Hsun Huang

Downloads last month
10

Models trained or fine-tuned on lianghsun/tw-judgment-gist