ShiZhi: A Chinese Lightweight Large Language Model for Court View Generation

TL;DR

This work tackles Criminal Court View Generation (CVG) by creating CCVG, a Chinese dataset with over 110K fact–court view pairs, and training ShiZhi, a 0.5B Chinese instruction-tuned LLM, on this data. The authors design a data-curation pipeline (section extraction, charge extraction, post filtering) and employ the Swift framework to fine-tune ShiZhi with a judge-like system prompt, achieving strong results on court view generation (ROUGE/ BLEU) and charge prediction (F1/accuracy). The key contributions are the domain-specific CCVG dataset and the lightweight ShiZhi model, which demonstrate that compact LLMs can produce legally coherent and fact-grounded court views when trained on high-quality domain data. This work provides a foundation for automated legal document generation and domain-focused AI in judicial contexts, with practical implications for efficiency and consistency in legal reasoning.

Abstract

Criminal Court View Generation (CVG) is a fundamental task in legal artificial intelligence, aiming to automatically generate the "Court View" section of a legal case document. Generating court views is challenging due to the diversity and complexity of case facts, and directly generating from raw facts may limit performance. In this paper, we present ShiZhi, the first large language model (LLM) specifically designed for court view generation. We construct a Chinese Court View Generation dataset, CCVG, of more than 110K cases, each containing fact descriptions paired with corresponding court views. Based on this dataset, ShiZhi achieving 70.00 ROUGE-1 and 67.85 BLEU-1 on court view generation, as well as 86.48\% accuracy with 92.75\% macro F1 on charge prediction. Experimental results demonstrate that even a small LLM can generate reasonable and legally coherent court views when trained on high-quality domain-specific data. Our model and dataset are available at \href{https://github.com/ZhitianHou/ShiZhi}{https://github.com/ZhitianHou/ShiZhi}.

Figures (4)

• Figure 1: The pipeline of our data curation and model training.
• Figure 2: The distribution of case occurrence years in the train and test sets.
• Figure 3: The distribution of charges in train and test sets.
• Figure 4: The length statistic of fact and court view in train and test sets.