If I understand your use case correctly, you are struggling with getting good performance out of your model using soft labels and in particular if the target is set to 0.5? If so, have a look at this post which shows that the cost of binary crossentropy is non-zero for this use case.