ChatLaw: Chinese Legal Model

Under the wave of ChatGPT, the continuous expansion and development of artificial intelligence have provided fertile soil for the spread of LLM. Currently, the fields of healthcare, education, and finance have gradually developed their own models, but there has been no significant progress in the legal field.

In order to promote open research on the application of LLM in law and other vertical fields, this project has open-source the Chinese legal model and provided a reasonable solution for the combination of LLM and knowledge base in legal scenarios.

The current open source versions of ChatLaw legal model for academic reference are Jiangziya-13B and Anima-33B. We use a large amount of original texts such as legal news, legal forums, laws, judicial interpretations, legal consultations, legal exam questions, and judgment documents to construct dialogue data.

The model based on Jiangziya-13B is the first version of the model. Thanks to Jiang Ziya's excellent Chinese language ability and our strict requirements for data cleaning and data augmentation processes, we perform well in logically simple legal tasks, but often perform poorly in complex logical legal reasoning tasks.

Subsequently, based on Anima-33B, we added training data and created ChatLaw-33B, which showed a significant improvement in logical reasoning ability. Therefore, it can be seen that large parameter Chinese LLM is crucial.

Our technical report is here: arXiv: ChatLaw

The version trained based on commercially available models will be used as the internal integration version for our subsequent products and is not open source to the outside world. You can try out the open source version of the model here