According to the sources, the bagging at each iteration should use a training set of a size of the original training set built up by randomly sampling, with repetition an original training set. So when the original training set contains, lets say, 8 samples, each new training set should also contain 8 samples, but some samples will be there multiple times, while others may be ommited.
Original set: A, B, C, D, E, F, G, H
Bag 1: A, C, A, G, E, C, H, H
Bag 2: C, A, H, F, D, H, D, B
Bag 3: F, E, D, D, A, E, B, A
The bagging metanode doesn't work like this - it uses a chunking loop, so trains the classsifers with a randomly chosen subsets of the training set without any bootstrapping.
...and WEKA nodes still don't work!