Plain file sparse data input format
-----------------------------------

Training and testing data are in separate files, same format.
Each row represents a case.

Row format and the way labels are presented depends on single-label/multi-label mode.

1. Single label mode
   -----------------

This is the most common situation, so the format here follows popular analogs like SVMlight. 
Row format is:

    <label> {<feature id>:<value>}*

Here <label> may take binary values 0/1 or 1/-1. The program only makes difference between the value 1,
which means positive, and any other, value which means negative.
Feature id is positive integer; value is double float.

2. Multi-label mode
   -----------------
In this case label values are stored separately from data, and data files include row names for reference
to labels.
Row format is:

    <row name> {<feature id>:<value>}*

Row name is a string; feature id is positive integer; value is double float.

Response values (labels) are stored in qrel files, same format as used by Vladimir's AdaptFilter program.
Qrel file refers to "documents" by their names. That's why the row name in the data file is important.
Qrel files for training and test may be the same or separate.
