ChatLaw-13B , this version is an academic demo version, based on the training of Jiang Ziya Ziya-LLaMA-13B-v1 , the Chinese performance is very good, but the legal question and answer with complex logic is not effective, and a model with larger parameters is needed to solve it .
ChatLaw-33B , this version is an academic demo version, based on Anima-33B training, the logical reasoning ability has been greatly improved, but because Anima's Chinese corpus is too small, English data often appear in questions and answers.
ChatLaw-Text2Vec uses a data set made of 930,000 judgment cases to train a similarity matching model based on BERT, which can match the user's question information with the corresponding legal articles, for example:
"What if the loan is not repaid?"
"Contract Law (1999-03-15): Article 206 The borrower shall repay the loan within the agreed time limit. There is no agreement on the loan period or the agreement is not clear, and it cannot be determined in accordance with the provisions of Article 61 of this law. Yes, the borrower can return it at any time; the lender can urge the borrower to return it within a reasonable period.”
The similarity between the two texts is calculated as 0.9960
Under the wave of ChatGPT, the continuous expansion and development of artificial intelligence has provided a fertile soil for the proliferation of LLM. At present, the fields of medical care, education, and finance have gradually developed their own models, but there has been no significant progress in the legal field.
In order to promote the open research of LLM in law and even other vertical applications, this project open-sources the Chinese legal model, and provides a reasonable solution in legal scenarios for the combination of LLM and knowledge base.
The current open-source versions of the ChatLaw legal model for academic reference are Jiang Ziya-13B and Anima-33B. We use a large number of original texts such as legal news, legal forums, legal articles, judicial interpretations, legal consultations, legal examination questions, and judgment documents. Construct dialogue data.
The model based on Jiang Ziya-13B is the first version of the model. Thanks to Jiang Ziya's excellent Chinese ability and our strict requirements for data cleaning and data enhancement, we perform well in legal tasks with simple logic, but complex logic Often perform poorly on legal reasoning tasks.
Then based on Anima-33B , we increased the training data and made ChatLaw-33B, and found that the logical reasoning ability has been greatly improved. It can be seen that the Chinese LLM with large parameters is very important.
Our technical report is here: arXiv: ChatLaw
The version trained based on the commercially available model will be used as the internal access version of our follow-up products. It is not open source to the outside world. You can try the open source version model here .
Visit Official Website