CSHaitao/LegalOne-4B
CSHaitao/LegalOne-4B is a 4 billion parameter, Qwen3-based large language model developed by CSHaitao, specifically trained for the Chinese legal domain. It utilizes a multi-stage training framework, including Plasticity-Adjusted Sampling for mid-training and Legal Agentic CoT Distillation for supervised fine-tuning, to enhance legal knowledge and reasoning. With a 40960 token context length, it excels in tasks like legal interpretation, case law reasoning, legal Q&A, and document drafting, outperforming general LLMs and existing legal models in its class.
Loading preview...
LegalOne-4B: Specialized LLM for Chinese Legal Reasoning
LegalOne-4B is a 4 billion parameter model from the LegalOne family, developed by CSHaitao, specifically designed for the Chinese legal domain. It addresses the limitations of general LLMs in handling the knowledge-intensive and structure-intensive nature of legal reasoning. The model is built on the Qwen3-4B-Base architecture and supports both Chinese and English.
Key Capabilities
- Enhanced Legal Knowledge: Incorporates a vast corpus of legal documents, including academic papers, court judgments, regulations, and legal consultations, ensuring deep understanding of legal concepts.
- Reliable Legal Reasoning: Employs a multi-stage training framework, including Plasticity-Adjusted Sampling (PAS) for mid-training and Legal Agentic CoT Distillation (LEAD) for supervised fine-tuning, to cultivate robust reasoning abilities.
- Agentic CoT Distillation: Utilizes a novel LEAD system that simulates professional legal workflows to generate high-quality, consistent reasoning trajectories, improving the model's "legal thinking" patterns.
- Strong Performance: Achieves competitive performance in tasks such as legal interpretation, case law reasoning, legal Q&A, and document drafting, often matching or surpassing larger general-purpose models on authoritative benchmarks like LexEval and JecQA.
- Multi-stage Reinforcement Learning: Applies curriculum learning to progressively shape reasoning capabilities from simple to complex tasks, guided by verifiable reward signals.
Good For
- Legal Research and Analysis: Assisting with in-depth analysis of legal cases, regulations, and academic literature.
- Automated Legal Q&A: Providing accurate and contextually relevant answers to legal queries.
- Document Drafting Support: Aiding in the generation of legal documents, summaries, and opinions.
- Legal Education and Training: Serving as a tool for understanding complex legal concepts and reasoning processes.
- Applications Requiring High Reliability: Ideal for systems where accuracy and consistency in legal outputs are paramount.