Does ByteCover's Triplet Loss Still Help When the Training Set Is Small? A 30-Run Ablation on Nested Subsets of SHS-100K-2025

ByteCover (Du et al., 2021) trains a ResNet-IBN encoder for cover song identification with a classification loss and a batch-hard triplet loss joined through a BNNeck, and reports that the combination outperforms classification alone on the full SHS100K training set. I ask whether the triplet term still helps when training data are scarce. I train two arms that share the encoder, GeM pooling, BNNeck and classifier and differ only in the presence of the triplet loss, on nested subsets of 50, 150 and 300 works from SHS-100K-2025, with five paired seeds per size (30 runs of 100 epochs each), and evaluate mean average precision (mAP) on Covers80. The paired difference in mAP (Arm B minus Arm A) is 0.015 ± 0.008 at 50 works, 0.011 ± 0.023 at 150 works and 0.023 ± 0.021 at 300 works (mean ± 95% paired t-interval), so the interval at 150 works includes zero. Every arm scores close to the chance level of 0.036 mAP and far below the published 0.906, and training length is confounded with dataset size. The observed gain is therefore small and should be read cautiously.