B3 set uses the channel with the same size of the input shape in the convolution layer. For example, (1,256,120) for dataset with 120 components, (1,256,90) for dataset with 90 components to avoid overfitting problem.
B4 set use the exact same channel with the benchmark.