In open-set object detection, low-quality datasets undermine model performance and training efficiency, leading to inaccurate category prediction and misaligned bounding box(bbox). Although these issues are well recognized, current research lacks effective methods for dataset quality assessment, while manual inspection remains costly and impractical for large-scale datasets. To address this, we propose a set of evaluation metrics specifically designed for open-set object detection, including category mislabeling rate, bbox offset rate, bbox redundancy rate, and—most notably—the image content utilization rate, to enhance assessment comprehensiveness. Building on these metrics, we develop an automated annotation enhancement method that leverages Multimodal Large Language Models (MLLMs) to correct and enrich dataset annotations. Experiments on Objects365, GQA, and RefCOCO demonstrate that our approach substantially improves both dataset quality and downstream detection performance. Specifically, our enhanced GQA dataset increases the zero-shot performance of OV-DINO on COCO2017_val by 8.6 mAP, while the refined RefCOCO annotations double training efficiency and yield a 22.8 mAP improvement. We release the improved GQA and RefCOCO datasets to support future research in this area.