Enhancing the accuracy of model evaluation accelerates the subsequent optimization process of large visual language models (LVLMs). However, evaluating the diversity of LVLM outputs is challenging. The long-form responses provided by LVLM are susceptible to interference from visual facts, which mainly manifests as factual inaccuracies stemming from not following the given image instructions. This problem makes the traditional Exact Match evaluation method no longer satisfied with measuring the correct expression of LVLM. To this end, this paper proposes KVQA-Eval, a novel evaluation benchmark aiming to reflect the discriminative ability of LVLM in the knowledge VQA domain. We propose a dataset (40k) called K-VQA, which improves the evaluation of answer correctness by incorporating human annotations. Furthermore, we train a scalable and deterministic FacGPT evaluator model based on GPT2. FacGPT not only achieves better correlation with human evaluators and GPT4 but also improves evaluation efficiency by 300%. More encouragingly, this evaluation model will help promote the development of automated evaluation tools due to their data safety, cost-effectiveness, and reproducibility. Our open-source code, dataset, and model at https://github.com/Siri-2001/FacGPT .