Trained model weights for "Lightweight attention mechanisms in breast ultrasound lesion segmentation: parameter cost, protocol choice, data leakage, and what the data can resolve"

Trained weights for eleven ablation configurations across three independent random splits (seeds 42, 142, 242), in training form. These are the files the evaluation path in the code record loads. MANIFEST.csv carries, for every file, the configuration name used in the paper, the key of its per-image table in the code record, its comparison family, size and SHA-256. CHECKSUMS.sha256 verifies the set. Two things will go wrong silently without reading the README. C7_F*.keras is the revised attention-gate implementation, not the original: the revised run wrote to the original checkpoint path and overwrote the earlier one, which is why the earlier implementation has no weights and is excluded from every paired analysis. Its per-image table is C7b. And the convolutional block attention variant cannot be loaded without attention_blocks.py from the code record; its custom layer is not registered for serialisation. Deployed-form models are not deposited; the parameter counts of both forms, and the counter that produces them, are in the code record.