Staged training method for large legal dialogue model

The invention relates to the technical field of large-scale language models, in particular to a staged training method for a legal dialogue large model, which comprises the following steps: determining training data used in different stages and preprocessing the data; a staged training method is used for improving the corpus quality and specialty of the answers of the model; a legal article retrieval module is introduced, and answers supported by evidences are generated; the staged training method for the large legal dialogue model has the beneficial effects that the professional ability of the large Chinese legal model is improved, so that the large Chinese legal model can give more professional, accurate and reliable answers when dealing with various questions in the legal field. The professional ability of the model in the legal field is improved through staged training adjustment and optimization and a targeted data set, meanwhile, a legal provision retrieval model is introduced, the final answer of the m