In order to achieve rapid prediction of structural seismic responses for emergency decision-making while addressing the high cost of acquiring target-specific data and the complexity of cross-scale prediction, a cross-scale transfer learning framework based on scaled-model data was used to investigate full-scale seismic response prediction for a 3D-printed concrete residential structure, and a hybrid GAE-Transformer-CNN model integrating graph attention mechanisms and positional encoding was developed, in which sensor spatial priors and structural topology were explicitly incorporated through a graph-aware module, long-range dependencies were captured by a Transformer, and local features were extracted by a CNN. The results show that after pre-training on the scaled model, fine-tuning was performed using only a small number of full-scale samples, and an average R² of 0.929 was achieved on an independent test set, with the maximum interstory drift ratio at critical locations being accurately predicted from ground-acceleration inputs and sparse sensor observations. It is concluded that scaled experimental data can be effectively leveraged to support data-efficient and accurate cross-scale seismic response prediction, thereby providing a practical and scalable methodology for structural response assessment.