As long as you're consistent, it makes no difference.
Either you put all the lower weights on left children and all the higher weights on right children, or vice-versa.
At the bottom line, left
and right
will only be variable names in your code, with no physical meaning whatsoever.
UPDATE:
If you are not consistent, then the resulting Huffman tree will not necessarily yield the best possible compression that can be achieved for the given input using the Huffman compression algorithm.