This paper presents MRSAudio, a large-scale multimodal recorded spatial audio dataset with refined annotations, designed for spatial audio generation and understanding tasks, along with its benchmarks.