Generate non-splittable test strings for worstcase benchmark #30
Conversation
hendrikvanantwerpen
force-pushed
the
unsplittable-test-strings
branch
from
7b94237
to
9931ddc
Compare
hendrikvanantwerpen
changed the base branch from
main
to
move-equivalence-tests
hendrikvanantwerpen
force-pushed
the
unsplittable-test-strings
branch
from
9931ddc
to
b21fc80
Compare
hendrikvanantwerpen
changed the base branch from
move-equivalence-tests
to
aneubeck/regex
| let text: String = ('\0'..char::MAX).filter(|c| !c.is_whitespace()).collect(); | ||
| for (name, tok, tiktoken, huggingface) in TOKENIZERS.iter() { | ||
| let text = create_test_string_with_predicate(&tok.bpe, 100000, |text| { | ||
| tok.split(text).nth(1).is_none() |
There was a problem hiding this comment.
this procedure has one downside. You pick ONE pattern with the first couple of bytes. But these different patterns probably have somewhat different performance. In particular the couple of apostrophe patterns are only up to four characters long :) So, if you are unlucky and pick one those, then the whole string construction will result in an endless loop...
There was a problem hiding this comment.
I don't think that can happen like that. The predicate is applied to the full text every time it is extended. So if you picked the token 're as the first token, it will be fine until you add anything else, because then it'll split after 're. The backtracking in create_test_string_with_predicate will give up adding anything after 're after a fixed number of tries, remove it, and pick another first token.
The worstcase benchmark used the continuous Unicode block as a test string. Now that we have the regexes available that are used for splitting, we can actually generate test strings that we know for sure cannot be split with the regex. This allows a better comparison of the underlying BPE performance. (Pre-tokenization still runs though and is part of the measurement!)