LawGPT: A Chinese Legal Knowledge-Enhanced Large Language Model

1 Introduction

• Proprietary LLMs accessed through APIs do not guarantee privacy for sensitive legal cases, while open-source models lack sufficient legal knowledge and downstream-task fit.The paper identifies these as two obstacles to practical legal use.Source: §1 · p. 1
• The authors frame an open-source model specifically designed for legal applications as a response to the privacy and performance obstacles.They describe LAWGPT as the first open-source model of this kind for Chinese legal applications.Source: §1 · p. 2
• LAWGPT surpasses the open-source LLaMA 7B model in major legal tasks.The authors present this as evidence for the model’s effectiveness.Source: §1 · p. 2
• LAWGPT is introduced as an open-source Chinese legal knowledge-enhanced model, with legal-oriented pre-training and legal supervised fine-tuning.The pre-training incorporates legal-domain knowledge; a knowledge-driven instruction dataset supports downstream-task adaptation.Source: §1 · p. 2

2 Related Work

• The section organizes prior work around general language models, legal language models, and legal benchmarks.These are the three areas the authors say the review covers.Source: §2 · p. 2

2.1 General Language Models

• Proprietary and open-source general LLMs are described as capable across downstream tasks, including legal applications.The paper notes that open-source models can be fine-tuned with targeted legal datasets.Source: §2.1 · p. 2

2.2 Legal Language Models

• Prior legal language models address tasks such as judgment prediction, retrieval, reading comprehension, and question answering, while newer approaches fine-tune legal corpora or use retrieval.The cited work also includes systems for legal reasoning and hallucination robustness.Source: §2.2 · pp. 2–3

2.3 Legal Benchmarks

• LawBench covers 20 legal tasks across knowledge memorization, understanding, and application; LAiW covers 14 tasks across three legal-capability levels.SimuCourt evaluates judicial analysis and decision-making.Source: §2.3 · p. 3

3 Methodology

• LAWGPT uses legal-oriented pre-training to add legal-domain knowledge to an open-source base model, followed by legal supervised fine-tuning for downstream legal tasks.The authors identify insufficient legal knowledge and task-specific training as challenges for open-source general LLMs.Source: §3 · p. 4

3.1 Legal-Oriented Pre-Training

• The legal pre-training corpus DLPT contains 500K legal documents from civil, criminal, and administrative law, and the model is trained autoregressively on legal text.The objective optimizes the base-model parameters using the corpus.Source: §3.1 · p. 4

3.2 Legal-Supervised Fine-Tuning

• The 300K-sample DLFT instruction dataset combines 200K open-source crime-task samples, 20K JEC-QA samples, and 80K legal QA samples refined with ChatGPT.The authors describe supervised fine-tuning on this dataset to adapt the pre-trained model to downstream tasks.Source: §3.2 · pp. 4–5

3.3 Inference of LAWGPT

• At inference, users wrap an instruction with the Alpaca testing template, tokenize it, and generate a response autoregressively with the fine-tuned model.The generated response is the model’s output for the supplied instruction.Source: §3.2 · p. 5, §3.3 · p. 5

3.2 Legal-Supervised Fine-Tuning

• Fine-tuning wraps DLFT instructions and outputs in the Alpaca training template and optimizes the pre-trained model on the output tokens.The paper describes this as the objective for obtaining the final LAWGPT parameters.Source: §3.2 · p. 5

4.1 Implementation Details

• Training uses a Chinese-Alpaca-Plus 7B base model, with LoRA-based legal-oriented pre-training on DLPT followed by legal supervised fine-tuning on DLFT.DLPT contains 500K legal pre-training corpus items and DLFT contains 30K legal-supervised corpus items; training used 8 NVIDIA V100 GPUs.Source: §4.1 · p. 5

4.2 Performance Evaluation

• In zero-shot evaluation across eight legal applications, LAWGPT outperforms LLaMA 7B on major tasks and achieves better average performance.Table 1 compares LAWGPT with proprietary models and 7B open-source models.Source: Table 1 · p. 6, §4.2 · p. 6
• Despite its data-privacy advantage, LAWGPT still has a significant performance gap with proprietary models.Source: §4.2 · p. 6

• LAWGPT is introduced as a Chinese legal knowledge-enhanced model designed for Chinese legal applications.Source: §5 · p. 6
• Legal-oriented pretraining incorporates legal-domain knowledge, while legal supervised fine-tuning enhances performance on downstream legal tasks.Source: §5 · p. 6
• Experiments report that LAWGPT outperforms open-source LLaMA 7B, and the authors hope it will inspire Chinese legal-AI research.Source: §5 · p. 6