A Long-Text Charge Prediction Method Based on the Prompt-Qwen-FIT Model: Integrating Few-Shot Learning and Incremental Adaptation Training

Charge prediction is a critical task in judicial AI, involving the determination of criminal charges through detailed analysis of case narratives. Existing methods often face high computational costs and limited data when processing lengthy and complex legal texts. To address these challenges, this paper introduces a novel model, Prompt-Qwen-FIT, designed to achieve efficient and accurate charge prediction in long case descriptions. First, the model extracts key defendant-related segments from the CAIL2018 partial dataset, creating a customized training dataset for fine-tuning. Then, it leverages LoRA technology to perform incremental adaptation and optimization on the Qwen model using the original data. Experimental results show that Prompt-Qwen-FIT achieves a classification accuracy of 94.36%, outperforming baseline models and demonstrating strong generalization capabilities and robustness in long-text classification tasks.