Abstract: Synthetic aperture radar (SAR) provides all-weather and day-and-night earth observation capabilities, holding significant application value in airport surveillance, maritime monitoring, and infrastructure inspection. Most existing publicly available SAR detection datasets employ horizontal bounding boxes for annotation, which often incorporate substantial background clutter within the boxes, thereby hindering the precise description of target features. Moreover, the vast majority of these datasets support only coarse-grained category detection, lacking fine-grained annotations such as aircraft or vessel types, and suffer from limited category diversity, making it difficult to establish a unified benchmark for multi-category testing. To address these challenges, the first publicly available version of the XuanJian series dataset, XuanJian-RSAR, which integrates eight publicly available SAR target datasets and forms a fine-grained rotated bounding box detection dataset containing 26 types of typical ground and sea targets, is constructed in this paper. To enable batch conversion from horizontal to rotated annotations, a semi-automatic annotation generation pipeline is proposed: the self-supervised symmetric branch of H2RBox-v2 is leveraged to predict target azimuths, rotated bounding boxes that satisfy projection matching constraints are achieved using a geometric constraint solver, and the annotation quality is ensured through comprehensive manual verification and correction. A systematic baseline experiment is conducted using 10 mainstream rotation detection algorithms, and quantitative detection indicators and visualization results for typical scenarios are provided, thereby offering a unified testing benchmark for the research of SAR fine-grained rotation detection algorithms and foundation remote sensing models.