A Large-Scale Coronal Mass Ejection Instance Segmentation Dataset with Catalog-Guided Selection and Two-Stage Human Refinement

CME-Seg-LASCO is a curated LASCO C2 coronal mass ejection (CME) instance segmentation dataset constructed from CDAW catalog-guided event selection and two-stage human refinement. The dataset contains 12,745 image-annotation pairs from 1,500 CME events observed between 2002 and 2023. Each sample includes a standardized 8-bit running-difference PNG image, a final refined LabelMe polygon annotation in JSON format, and a corresponding rasterized binary mask PNG. Files are organized hierarchically by year and event, preserving temporal order for sequence-based learning. The dataset is designed to support benchmarking and training of single-frame and sequence-based CME segmentation and tracking models, and includes events spanning narrow, limb, partial-halo, and full-halo CME categories.