Thanks to a FB comment by Jose Carlos, I have discovered this really interesting dataset. Oscar Celma, Ph.D, Chief Innovation Officer @ Barcelona Music & Audio Technologies, has written a very interesting thesis on Music Recommendation, with the plus of making his dataset available to the community.
This dataset contains tuples (for ~360,000 users) collected from Last.fm API, using the user.getTopArtists() method, and it is ~543Mb. The data is made available for non-commercial use. As it essentially features <user, artist, plays> tuples, it is very interesting for testing a number of collaborative recommendation techniques and functions (recommending music & artists). However, I miss it was date-tagged, in order to examin how a system would evolve regarding recommendations and feedback.
It contains 17,562,018 tuples, beware if you are going to use non-sparse representations.