There are two sets of SRL data in the SNoW format. One is "argument-identification", which is an intermediate stage that filters out some argument candidates. The label of each example is either 1 (positive) or 2 (negative). The other is "argument-classification", which is the final stage that decides the labels of each argument candidates. It has 45 labels, and label 45 means "null" or "not a candidate".
In both data sets, the features represent some syntactic information extracted from Charniak's parser. The file names indicate whether the data is generated from treebank sections 02 to 21, or section 23. Also, in each folder, the .res file record the predictions of the SNoW classifier. Note that we didn't use the full 45 classes when we trained the argument classifier; therefore, the corresponding .res file only has the predictions for some classes.