For class prediction, you can use 'golub_*.txt', which have normalized expression values for each gene and each sample.
'golub_training.txt' is training data. Each row represents each gene, and each column represents each sample. (Training data has total 38 samples.)
'golub_test.txt' is test data. Each row represents each gene, and each column represents each sample. (Test data has total 34 samples.)
'golub_truth.txt' is the ground truth for total 72 samples (training 38 samples + test 34 samples).