1. You should be aware that <<ENDS is dealt like <<"ENDS" i.e. allows variable interpolation. Better write <<'ENDS' to avoid this.

2. you are free to use multiple here-docs in the same line, so if the number of blocks is too small for concerns about being DRY, you can write

use Data::Dump;

my %desc;
init_desc();

dd \%desc;

sub init_desc {
    @desc{ONE,TWO}= (<<'__ONE__',<<'__TWO__');
one
__ONE__
two
__TWO__
}
[download]

3. Please note that I've put the population part away into a sub init_desc() , like this you can have multiple of such initializations hidden at the end of your code.

4. Aforementioned solution isn't as DRY as you wanted, but actually your split-solution wasn't too bad, though your grep to ignore the first line is dangerous:

my %desc = init_desc();

dd \%desc;

sub init_desc {
    (undef,my %hash) =                                      #  ignore 
+first line
      split /^ \[ (\w+) \] \s* $/xm,  <<'__ENDS__';
[ONE]
one

[TWO]
two
__ENDS__

    return %hash;
}
[download]

(potential trimming of leading and trailing "\n" is left as an exercise).

5. please note that the last approach can also be used to parse a slurped __DATA__ section.