SHerZH/IMLJP

MMSI Dataset

⚠️ Current status
The current release of SHerZH/IMLJP is a small preview subset of the full dataset.
It is intended only for visualizing the data structure and format, and for small-scale experiments.
A larger / full version of the dataset may be released in the future after further cleaning, annotation checking, and privacy review.

This repository hosts a processed dataset used in the paper:

"Explainable Multidefendant Judgment Prediction Enhanced by Judicial Logic Based on Large Language Models"

The dataset consists of Chinese criminal judgments involving multiple defendants, with:

• A condensed fact description (FD),
• A court reasoning segment (CV),
• And per-defendant structured labels (role, prison term, probation), stored as a list of defendants.

The current subset only includes cases whose main offense is intentional injury (Article 234 of the PRC Criminal Law).

Data source and privacy

• Raw judgments are collected from China Judgments Online (中国裁判文书网), which provides publicly available court decisions.
• We only release processed and anonymized data:
  • Personal names are masked (e.g. [张某], [王某1]);
  • Highly sensitive information (ID numbers, phone numbers, exact addresses, etc.) is removed or obfuscated;
  • Only fields necessary for modeling (e.g. FD, CV, and per-defendant sentencing labels) are preserved.

The dataset is intended solely for research and educational purposes.
Users must not attempt to re-identify any individual or link cases back to specific natural persons.

🇨🇳 简短中文说明：
本数据集基于中国裁判文书网上公开的、以故意伤害罪为主要罪名的刑事判决书构建，
对当事人姓名等信息进行了脱敏处理，仅保留建模所需的“事实摘要（FD）”、“裁判说理（CV）”及按被告人划分的量刑标签。
当前版本为小规模样例集，主要用于展示数据结构和辅助复现。

---

1. Repository & Code

• Code implementation for the paper (model, training, experiments) is provided in the GitHub repository:

  👉 GitHub: Logic-Guided Multistage Inference for Explainable Multidefendant Judgment Prediction

The Hugging Face dataset SHerZH/IMLJP is mainly intended to:

1. Provide a public example of the data format (e.g., FD, CV, defendants list with per-defendant labels).
2. Help users understand how to construct their own datasets in the same structure based on publicly available judgments (e.g., from China Judgments Online).
3. Support lightweight tests and visualization of models and preprocessing pipelines.

In future, we plan to release a larger / full dataset version (with the same structure and the same crime category: intentional injury) once all ethical, privacy, and legal considerations are carefully addressed.