fix: preserve all token IDs when multiple BPE tokens decode to same string#1833
Merged
RobinPicard merged 2 commits intodottxt-ai:mainfrom Mar 26, 2026
Merged
Conversation
RobinPicard
requested changes
Mar 17, 2026
Contributor
There was a problem hiding this comment.
Hi, thanks for opening a PR! Another PR has already been merged to fix this issue though (#1831). What you could do is to rebase on main to get the fix from the other PR, and then have this PR be about having more thorough testing of the feature since you have more tests than there are on main.
tests/backends/test_outlines_core.py
Outdated
| assert "name" in response | ||
|
|
||
|
|
||
| def test_create_vocabulary_preserves_duplicate_token_ids(): |
Contributor
There was a problem hiding this comment.
That's a good test to have, but the content/description below does not include different tokens that decode to the same string!
Author
|
Thanks for the review, Robin! I'll rebase on main to pick up the existing fix and refocus this PR on the improved test coverage. Good catch on the first test's description not matching its content — I'll fix that as well. |
alvinttang
force-pushed
the
fix/vocab-token-id-overwrite
branch
from
de2c485 to
2aff36b
Compare
Contributor
|
Tests are failing! |
…tring In `create_outlines_core_vocabulary`, the assignment `formatted_vocab[token_as_str] = [token_id]` overwrites previous entries when multiple BPE tokens decode to the same string via `token_to_str`. For Mistral-7B, this silently drops 17.2% of token IDs. Replace with `formatted_vocab.setdefault(token_as_str, []).append(token_id)` to accumulate all token IDs for each decoded string. Fixes dottxt-ai#1830 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
RobinPicard
force-pushed
the
fix/vocab-token-id-overwrite
branch
from
2aff36b to
05d198e
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Fixes #1830
Bug: In
OutlinesCoreBackend.create_outlines_core_vocabulary(), the line:overwrites previous entries when multiple BPE tokens decode to the same string via
token_to_str(). For Mistral-7B, this silently drops ~17.2% of token IDs from the vocabulary, causing the structured generation FSM to miss valid token transitions.Fix: Replace the assignment with:
This accumulates all token IDs that map to the same decoded string, preserving the complete vocabulary for the
outlines_core.Vocabularyconstructor.Changes
outlines/backends/outlines_core.py: One-line fix increate_outlines_core_vocabularytests/backends/test_outlines_core.py: Two new unit tests:test_create_vocabulary_preserves_duplicate_token_ids— verifies basic vocabulary building with sentencepiece-style tokenstest_create_vocabulary_duplicate_decoded_strings— verifies that whentoken_to_strmaps multiple tokens to the exact same string, all IDs are accumulated (the core regression case)Test plan
pytest tests/backends/test_outlines_core.py::test_create_vocabulary_preserves_duplicate_token_ids tests/backends/test_outlines_core.py::test_create_vocabulary_duplicate_decoded_strings)pytest tests/backends/)