Extremely large-scale multiple-input-multiple-output (XL-MIMO) is envisioned as a key enabler for future six-generation (6G) networks. Nevertheless, the pronounced expansion in both antenna aperture and operational frequency mandates the consideration of near-field conditions. This shift complicates traditional far-field beam training designs, which struggle with near-field spherical-wave-based modeling, thereby compromising channel capacity. Specifically, in near-field systems, unfavorable computational complexity and pilot overheads become almost inevitable, given that beam training codebooks are coupled with both angles and distances of scatterers in significant propagation paths. Moreover, the power leakage effect in the near-field system can erode beamforming gains if not addressed. To counter these challenges, we introduce a novel two-stage learning-based beam training protocol that independently manages angles and distances. Building on this foundation, we present a hierarchical codebook design with modular narrow codeword units termed hierarchical codebook design via accumulation (HCD-A) to alleviate the power leakage problem. Our simulation results underscore the superior performance of our proposed solutions at different signal-to-noise ratio (SNR).