Skip to content

fix(tokenizer): extend decoding fix to cover Mistral no-spaces bug #2285

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
eeshsaxena wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix(tokenizer): extend decoding fix to cover Mistral no-spaces bug #2285

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 4 additions & 2 deletions litgpt/tokenizer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -68,12 +68,14 @@ def __init__(self, checkpoint_dir: Path | str) -> None:
raise NotImplementedError

# NOTE: A temporary fix until it's resolved on Tokenizers side.
# LlaMA tokenizer strips leading spaces if to decode a single token at a time.
# LLaMA and Mistral tokenizers strip leading spaces when decoding a single token at a time,
# because both are backed by SentencePiece which encodes word boundaries with a prefix space (▁).
# https://github.com/huggingface/transformers/issues/31643
self.apply_decoding_fix = None
if (config_path := checkpoint_dir / "tokenizer_config.json").is_file():
with open(config_path, encoding="utf-8") as fp:
self.apply_decoding_fix = "LlamaTokenizer" in json.load(fp)["tokenizer_class"]
tokenizer_class = json.load(fp).get("tokenizer_class", "")
self.apply_decoding_fix = "LlamaTokenizer" in tokenizer_class or "MistralTokenizer" in tokenizer_class

@property
def vocab_size(self) -> int:
Expand Down