Applying Merges In Rank Order

A BPE tokenizer was trained and produced this merge table. Merges are applied to a word in rank order, rank 1 first, and each merge is applied everywhere it matches before the next one is considered.

rank	merge	new symbol
1	`u` + `g`	`ug`
2	`h` + `ug`	`hug`
3	`p` + `u`	`pu`
4	`n` + `</w>`	`n</w>`

The word pug never appeared in training. At inference it is first split into characters, p u g </w>. How is it tokenized?

pug </w>

p ug </w>

pu g </w>

p u g </w>