Architecture and training workflow of MTLforGalSpecZ. The model comprises three main components: (1) a shared feature extraction module with multiple convolutional blocks to process input spectra; (2) a channel attention mechanism to enhance task-relevant features; and (3) two task-specific output heads—one for spectral classification of galaxies and the other for gas-phase metallicity regression. Training is guided by a joint loss function that combines focal loss for classification and masked mean absolute error (MAE) loss for regression, enabling end-to-end multi-task optimisation.