Accuarcy of ZeroR with k-fold Cross Validation

[Akshay Nayak]

Upon trying k fold CV with ZeroR classifier on a number of datasets, I am observing a few things:

1)If the number of instances in each class is equal, then the accuracy of ZeroR for k fold CV on n observations varies with the value of k with it approaching 0 as k approaches n(which is obvious considering a majority in the test set cannot be a majority in the training set). However my doubt is regarding the other values of k. The accuracy seems to change with different values of k and I cannot get a fixed value for determining the baseline accuracy.

2) Regarding the accuracy of 10 fold CV on the class 'play' in the weather nominal dataset - For the sake of understanding how stratified CV works, I tried to divide 'yes'(9) and 'no'(5) in the following manner.
Fold 1 - Yes and No
Fold 2 - Yes
Fold 3 - Yes
Fold 4 - Yes
Fold 5 - Yes and No
Fold 6 - No
Fold 7 - Yes and No
Fold 8 - Yes
Fold 9 - Yes
Fold 10 - Yes and No.
After computing the probabilities for each fold and averaging them, I am getting an accuracy of 70%. I know that Weka invokes the learning algorithm an 11th time on the entire data set giving a final accuracy of 64.285% but how do we get that accuracy after getting a 70% accuracy for the first 10 folds?