fix edge case for qwen3 data processing #626
Merged
Conversation
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
RobotSail
force-pushed
the
fix-data-processing
branch
from
d9a21d8 to
79c2f2e
Compare
mergify
bot
added
documentation
RobotSail
force-pushed
the
fix-data-processing
branch
7 times, most recently
from
765eb34 to
414c296
Compare
Contributor
|
@RobotSail , we are working on CI. In the meantime please just run the Large e2e job manually on this PR when its ready for review |
|
E2E (NVIDIA L40S x4) (python 3.11) workflow launched on this PR: View run |
|
e2e workflow succeeded on this PR: View run, congrats! |
Contributor
|
@Mergifyio rebase |
…_content` fields Signed-off-by: Oleg S <[email protected]>
Signed-off-by: Oleg S <[email protected]>
…: false` on samples functions as expected Signed-off-by: Oleg S <[email protected]>
Contributor
✅ Branch has been successfully rebased |
cdoern
force-pushed
the
fix-data-processing
branch
from
968ba35 to
a05d1e4
Compare
cdoern
approved these changes
Jun 30, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 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Outdated suggestions cannot be applied.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.
With Qwen3, there's an edge case which can result in the unmask/mask logic breaking during data processing.
Root Cause: The error occurs specifically when using the Qwen/Qwen3-32B tokenizer, not with Qwen/Qwen2.5-32B-Instruct. The problematic sample contains multiple tags in the assistant's
response.
Issue Location: The error occurs in data_process.py:555 in the unmask_messages function, where it encounters an <|UNMASK_END|> token while not in an unmasking state.
Key Findings:
The Problem:
The Qwen/Qwen3-32B model's chat template is processing the <|UNMASK_BEGIN|> and <|UNMASK_END|> tokens in a way that causes them to appear out of order or in an unexpected state, leading to
the algorithm encountering an <|UNMASK_END|> token when it's not actively unmasking.
This is likely due to differences in how the chat templates of Qwen2.5 vs Qwen3 handle special tokens, particularly when there are multiple special tokens or complex content like the
tags present in the assistant's response.
Signed-off-by: Oleg S [email protected]