This is an old corpus used for email spam detection.

Some messages are known to be unwanted mass emails ("spam");
the rest are known to be genuine ("gen").

The GenSpam corpus was prepared by Medlock (2006):
   https://www.researchgate.net/publication/220271836_An_Adaptive_Semi-Structured_Language_Model_Approach_to_Spam_Filtering_on_a_New_Corpus

Here is a quote from that paper:

   We source the Adaptation and Test sets from the contents of two users
   inboxes, collected over a number of months (Nov 2002–June 2003),
   retaining both spam and genuine messages. We take this approach rather
   than simply extracting a test set from the corpus as a whole, so that
   the test set represents a real-world spam filtering instance. The 600
   messages making up the adaptation [development] set are randomly
   extracted from the same source as the test set, facilitating
   experimentation into the behaviour of the classifier given a small set
   of highly relevant samples and a large background corpus.

For this homework, we have provided training sets of different sizes: e.g.,
   gen            spam	     
   gen-times2	  spam-times2
   gen-times4	  spam-times4
   gen-times8	  spam-times8

Each of these training files contains many messages -- one message per
line.  gen-times8 and spam-times8 are the complete sets of genuine and
spam training messages, but it may be convenient to start with smaller
datasets such as gen and spam.