Open
Conversation
Contributor
|
These are default values and not the recommend values (especially for newer models). Please pass empty string |
Author
|
Thanks for the info. We have talked to a few labs and none of them are aware of this default value. As a result, many of them often fail to reproduce paper results as they just accept the default. Since it is quite safe to assume that most people are not aware of "###" being the default, we suggest changing the default to make it easier for the researchers! |
Will anything break if the default value is left empty? |
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
We identified a bug in the parser.py file, where the default stop token for model generation is "###":
This will cutoff model response prematurely. For example, in our evaluations, we find that Qwen3-8B likes to output "### Solution code" so the solution code is discarded.
Currently, this feature is not disabled for vllm runner, which is the default entry point for most open source models.
We changed the default flag to None and also handled the none type in subsequent logic.