System load pressure testing method for large language model

• Main Authors: JI SHOULING, XU YIXIAO, LIU XIAOLEI, DING KANGYI, XIN BANGZHOU, PU YUWEN
• Format: Patent
• Language: Chinese; English
• Published: 14.05.2024
• Subjects: CALCULATING; COMPUTING; COUNTING; ELECTRIC DIGITAL DATA PROCESSING; PHYSICS

The invention relates to the field of large language models, and provides a system load pressure testing method for a large language model. The method aims to overcome the defects that an existing sponge test sample generation method is low in generation efficiency, limited in test effect and the like. According to the main scheme, the method comprises the steps of sponge test sample generation, wherein a sponge test sample is generated in a white box gradient optimization mode for a local large language model, and a pressure test sample data set is constructed; and a large language model pressure test step: carrying out system load pressure test on the large language model to be tested according to the pressure test sample data set and generating a test result. According to the method, the generation of the white-box sponge test sample is realized by introducing the Gumbel-Softmax technology, the generation efficiency is improved, and the activation degree of the expert model is brought into an optimization