Along with the rapid development of large language models, evaluation datasets have also attracted research attention. They are proposed to evaluate models during text generation to ensure reliability in use. In evaluation tasks, factuality is always the primary criterion. In this study, we introduce a dataset with 4,023 QA pairs to evaluate the authenticity of LLMs in the application of Traditional Vietnamese Medicine (TVM). The dataset consists of short QA pairs that help assess the model quickly and accurately. In addition, we propose an effective process for creating similar factuality evaluation datasets with comprehensive quality control. We also evaluate 14 LLMs, divided into three groups with different parameter counts. The results show that the F1-score is linearly related to the number of model parameters, indicating that the dataset meets the requirements for a competency assessment test. In the tests, the Qwen models perform best across the groups. The fine-tuned Vietnamese model, Vistral-7B-Chat, with only 7B parameters, delivers quite impressive results, with an F1-score of 26.3, compared to the group average of 19.6 and the Qwen3:8b model, which has the highest score of 27.7.