Skip to content

Fix tokenizer typos, add newlines

Alexandru-Mihai GHERGHESCU requested to merge fix/tokenizer into main

Fix some issues in the tokenizer.

Mainly, fix a problem where newlines weren't present in the tokenizer model. This meant any whitespace was silently deleted and newlines weren't a thing. This could introduce issues for datasets such as wikitext103, where newlines delimited titles and actual text.

Re-train the wikitext tokenizers with newlines.

Merge request reports