Skip to content

Conversation

ArthurZucker
Copy link
Collaborator

@ArthurZucker ArthurZucker commented Jul 12, 2024

Very draft for now:

  • handle potential collisions

fixes #1473

@HuggingFaceDocBuilderDev

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

@pritam-dey3
Copy link

We really need this feature! Is it tracked anywhere else?

@ArthurZucker ArthurZucker reopened this Aug 29, 2024
@ArthurZucker
Copy link
Collaborator Author

oups sorry, yeah here is the good place will come back in a bit!

@ArthurZucker
Copy link
Collaborator Author

Okay, so this PR works for assigning, but the issue is that for Unigram based models if the world is not removed from the internal lattice, it can still be "encoded". Might be an issue.
We are also updating the added_tokens layer on top of vocab

@ArthurZucker
Copy link
Collaborator Author

Related to #1437 that does not work for unigram as if the token is in the vocabulary, even if it is special, it's still added to the unigram algorithm!

@jp1924
Copy link

jp1924 commented Jan 21, 2025

hello @ArthurZucker
Is this work simply delayed, or has it been stopped due to a critical issue?

related issue: #1473, #31475

@ArthurZucker
Copy link
Collaborator Author

there is kind of a critical issue which is that a general solution is hard to find given how unigram works. Once backtracking BPE is merged i'll have another look

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
Projects
None yet
Development

Successfully merging this pull request may close these issues.

Assign <unusedXX> tokens with special_tokens without growing vocab size
4 participants