Keda Xunfei announces the publication of a large-scale domain question and answer model training patent, which can save on labor costs.

Tianyancha's property clues information shows that on May 9th, the patent for "Domain Question-Answer Large Model Training and Question-Answer Method, Related Equipment and Program Products" applied by iFlytek Co., Ltd was published. The abstract shows that this application introduced a referee model, which can give preference scores to the input answers and evaluate the quality of the answers. Based on the domain knowledge base, domain question-answer data is extracted. During the iterative training of the initial large model, the referee large model scores the first answer corresponding to the domain question generated by the initial large model, and the second answer in the domain question-answer data, respectively. Based on the scoring results, answers that meet the preference requirements are selected to form target training data along with the domain questions. By using the target training data to train the initial large model, the final domain question-answer large model is obtained, which uses high-quality domain question-answer training data for training. This application does not require manual verification of the extracted domain question-answer data to obtain high-quality target training data, saving on manual costs and improving training efficiency.