Existing cross-modal person re-identification datasets still lack full-scale benchmarks for sketch-to-RGB person retrieval. To address this problem, this paper investigates sketch-to-RGB cross-modal person re-identification. First, this paper constructs a full-scale Sketch-Market1501 dataset based on Market-1501, containing 16,304 generated sketch-style images, including 12,936 images corresponding to the training subset and 3368 query images, and covering all 1501 person identities in Market-1501. By retaining the original annotations and official data partitions, a standardized sketch-to-RGB retrieval protocol is established, in which 3368 generated sketch-style images serve as queries and 15,913 original real-world RGB images form the gallery. Using OSNet-AIN, which incorporates style normalization, as the baseline, this paper systematically investigates the effect of different training-sample mixing ratios. The experimental results show that $p=0.75$ yields a Rank-1 result of 61.2% and is subsequently used in the attention-module and loss-function studies. At the high-level feature stage, an Efficient Channel Attention (ECA) module is embedded to recalibrate high-level channel responses and improve cross-modal feature discrimination. Furthermore, a joint loss combining Softmax loss, Triplet loss, and Focal loss is adopted to optimize feature discriminability. The experimental results indicate that the constructed dataset supports sketch-to-RGB cross-modal training. The final model achieves 43.3% mAP and 63.0% Rank-1, representing improvements of 3.6 percentage points in mAP and 1.8 percentage points in Rank-1 over the OSNet-AIN baseline under the same $p=0.75$ setting, thereby providing a benchmark and reproducible baseline for future research on sketch-based person re-identification. The proposed Sketch-Market1501 dataset will be publicly available after publication of the paper.