Heck, are you STILL using the old VGR code?? There is a better version that the guys fixed up and saved a lot of memory (search for Adventure+Optimized).
Anyway, a source code will contain the labels in a table for the various rooms. The labels in the table point to the data for a particular room...let's call it LFE75 and assume that it holds the above data at that address. So the program grabs the first byte there, $CE, and drops it into a ram location...let's call it R1. Then the program grabs the second value, $FB, and plunks it into an adjacent ram location - $R2. Now when the kernal needs the pixel data for the room, it just uses the Y register as an offset to point to the data line that it is currently drawing, and LDA ($R1),Y will grab the byte that it needs.
In the optimized source code, that CE and FB have both been changed into labels...so if you add program instructions or data tables, etc., it will point to the correct address when it's compiled.
Again, this is not data that you (as the programmer) even needs to worry about. You just need to make sure that the data line is containing the correct < and > labels in it. In the optimized code, you'll see lines like...
.byte
Yellow_Castle,$00,$00,$00...etc.
(actual data not used because I was too lazy to look them up) 
Then when Dasm compiles the code for you, it checks it's list of indexes when it reaches that line...
"Hmm...The tag Yellow_Castle is now at address $FD00..."
...and it puts the correct values into the line...
.byte $00,$FD,$00,$00,$00...etc.