Each "point" is a vector of values. Are you saying that each element of those vectors is derived from 1 of 5 states? e.g.,
[A A C C B; %point 1
A C E E B; %point 2
D A B A C] %point 3
And your goal is to classify these points into precisely two classes total?
If that is the case, then just go ahead and feed your data in to svm, splitting your overall data between training and test however is appropriate (e.g., sometimes half and half are used.)
If, however, you mean something quite different, that you have data points that you want to classify in to 5 different states, then svm is not appropriate for that. svm is a binary classifier. Some people have worked on extending svm to more classes, but the results are pretty time-consuming. If I recall, I did find one paper with an approach whose big-O had a logN in one place where the others had N, but I seem to recall that needed a lot of memory.
