This dataset contains the full experimental results, pre-trained model weights, and supporting files for LASD-Net (Lightweight Attention-Enhanced Siamese Deep Neural Network with Harmonic Subband Decomposition and Meta-Learning), a novel architecture for SSVEP-based Brain-Computer Interface (BCI) classification. Conducted as a Diploma IV (D4) Final Project at the Department of Informatics Engineering, Universitas Logistik dan Bisnis Internasional (ULBI), 2026. Contents: Per-experiment JSON result files (Experiments A–E, Ablation Study, EEGNet Zero-Shot, Statistical Tests), visualisation figures (PNG), pre-trained PyTorch model weights (global_model_lasd.pt), and model card with architecture details. Key results: LASD-Net achieves 20.3% cross-subject accuracy (LOSO-CV, 20 subjects), exceeding its intra-subject accuracy of 15.7% — a +4.5 pp inversion confirming MAML-driven calibration-free generalisation. Full HSA yields +7.3 pp over no-attention baseline (Wilcoxon p < 0.001, r = 0.88). Model size: 17,903 parameters, 15.55 ms CPU inference latency. Dataset used: Benchmark Tsinghua SSVEP Dataset (Wang et al., 2017). Raw EEG data is NOT included