GBMT-SLID is a global bimodal–bitemporal landslide inventory dataset constructed from co-registered Sentinel-2 pre- and post-event imagery and Copernicus DEM products across 29 diverse regions worldwide. To ensure consistency and analytical readiness, all optical and topographic layers were upscaled to 10 m resolution, harmonized, and processed using the IMAD algorithm to generate change maps, while expert landslide inventories were rasterized to aligned 10 m masks. Sentinel-2 spectral bands and spectral information features were stacked into 11-channel pre/post image pairs, and DEM-derived topographic factors formed 6-channel topographic inputs. All data were normalized and tiled into 256×256 patches using a 64-pixel stride, with strict quality control excluding patches with insufficient landslide content or excessive NoData, followed by KNN-based imputation and visual alignment checks, resulting in 3,772 high-quality multimodal patches. A two-stage enhancement pipeline diffusion-driven domain adaptation and region-aware augmentation—was applied before partitioning the dataset into 24 “seen” regions for training/validation and five “unseen” regions (Colombia, DRC, Uganda, Myanmar, Philippines) for generalization assessment. GBMT-SLID is designed as a comprehensive benchmark for advancing landslide segmentation, multimodal fusion, and bitemporal analysis in remote sensing