The Tianyancha App shows that Tencent Technology (Shenzhen) Co., Ltd. has recently applied for a patent titled "Training Method, Device, Computer Equipment, and Storage Medium for Large Language Models." The summary of this patent reveals that the method introduces first and second summary texts to provide the model with more learnable information during the training process of the large language model.
According to the patent description, the first and second summary texts contain different amounts of information, with the first summary text including both correct and incorrect statements. By performing comparative learning on these two different summaries of the same text and distinguishing between the correct and incorrect statements, the method effectively avoids issues such as model overfitting and inaccurate generation that may arise from a singular summary text.
The innovation of this method lies in enhancing the model's generalization performance and significantly improving its accuracy. By introducing diverse summary text content, Tencent's training method brings more efficient and precise enhancements to the training process of large language models.
