NB: Using the same precompiled regexes each time.

use v5.12;
use warnings;
use Data::Dump;

my $re_start = qr{(?<start>\w+)\s*};

my $re_repeat = qr{\s*(?<repeat>\d+)\s*};

my $str = 'a 1 2 3 b 4 5 6';

while ( $str =~ /\G$re_start/g ) {
      print "$+{start}: "; 
      while ( $str =~ /\G$re_repeat/gc ) {
            print "$+{repeat} ";
      }
      print "\n";
}


my @res;
$str =~ /
            (?:
                $re_start
                (?{ push @res, $+{start} })

                (?:
                    $re_repeat
                    (?{ push @res, $+{repeat} })
                )*
            )*
        /gx;

dd \@res;
[download]

a: 1 2 3 
b: 4 5 6 
["a", 1, 2, 3, "b", 4, 5, 6]
[download]

edit

• https://perldoc.perl.org/perlretut#Global-matching
• https://perldoc.perl.org/perlretut#A-bit-of-magic:-executing-Perl-code-in-a-regular-expression

to meet these explicit requirements, here an altered version.

NB: how similar those solutions look like in the end...

use v5.12;
use warnings;
use Data::Dump;

my $str = 'a 1 2 3 b 4 5 6';

my $re_start  = qr{(?<start>\w+)\s*};
my $re_repeat = qr{\s*(?<repeat>\d+)\s*};


# --- approach 1 with \G and /gc 
my @res1;
while ( $str =~ /\G$re_start/g ) {
      push @res1, [$+{start}]; 
      while ($str =~ /\G$re_repeat/gc) {
            push @{ $res1[-1] }, $+{repeat};
      }
}
dd \@res1;

# --- approach 2 with embedded (?{CODE})
my @res2;
$str =~ /
            (?:
                $re_start
                (?{ push @res2, [$+{start}] })

                (?:
                    $re_repeat
                    (?{ push @{ $res2[-1] }, $+{repeat} })
                )*
            )*
        /gx;
dd \@res2;
[download]

output

[["a", 1, 2, 3], ["b", 4, 5, 6]]
[["a", 1, 2, 3], ["b", 4, 5, 6]]
[download]

Cheers Rolf
_{(addicted to the Perl Programming Language :)
Wikisyntax for the Monastery}