BAI-Law-13B Law-LLM Development

Shanghai Jiao Tong University Artificial Intelligence Institute & Intelligent Law Institute & Baidu Intelligent Cloud

2023-09 - now

Based on the open-source LLM Llama-2, we performed domain post-pretraining and further scenario-supervised fine-tuning with a large amount of data such as judicial documents, legal documents, and law books. In the third-party legal comprehensive evaluation benchmark LawBench test, BAI-Law-13B outperforms all the current publicly available Chinese generalized LLMs and domain fine-tuned LLMs.

I was personally involved in the construction of the training data and led the training data screening. I performed the LawBench evaluation and realized further performance improvements in conjunction with Retrieval-Augmented Generation.