feat: add VLMEvalKit-compatible Qwen task variants for MMMU and MMStar#1021
Merged
feat: add VLMEvalKit-compatible Qwen task variants for MMMU and MMStar#1021
Conversation
Add new task variants that use VLMEvalKit-style prompt formatting:
- mmmu_val_qwen: Uses 'Question: {q}' prefix and 'Answer with the option letter only.' suffix
- mmstar_qwen: Uses same VLMEvalKit-compatible prompt structure
These variants help users reproduce benchmark scores closer to official Qwen
results reported in VLMEvalKit evaluations.
Usage:
python -m lmms_eval --model qwen2_5_vl --tasks mmmu_val_qwen,mmstar_qwen ...
Addresses score reproduction gaps reported in Issues #935, #932, #881, #901
kcz358
reviewed
Jan 23, 2026
kcz358
reviewed
Jan 23, 2026
Collaborator
There was a problem hiding this comment.
A small mismatch to the open ended propmt but the same motivation compare to #929. Seems this version more make sense to me?
|
Hi there, As I used lmms-eval main branch #714f4fed7, for Qwen3-VL-4B-Instruct the results are as follows: 
Still 0.6216 is a 7% gap compared to the official report 0.698 with respect to MMStar metric. Is there any suggestion or any further merge can resolve this issue ? Best, |
This was referenced Feb 9, 2026
Luodian
added a commit
that referenced
this pull request
Feb 28, 2026
#1021) Add new task variants that use VLMEvalKit-style prompt formatting: - mmmu_val_qwen: Uses 'Question: {q}' prefix and 'Answer with the option letter only.' suffix - mmstar_qwen: Uses same VLMEvalKit-compatible prompt structure These variants help users reproduce benchmark scores closer to official Qwen results reported in VLMEvalKit evaluations. Usage: python -m lmms_eval --model qwen2_5_vl --tasks mmmu_val_qwen,mmstar_qwen ... Addresses score reproduction gaps reported in Issues #935, #932, #881, #901
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.
Summary
Adds VLMEvalKit-compatible task variants for Qwen models to address the 10-20% score gaps reported between lmms-eval and VLMEvalKit.
Problem
Users reported significant score differences when evaluating Qwen models:
Root cause: Different prompt formatting between the two frameworks:
"Options: A: option1\nB: option2""Question: ...\nOptions:\nA. option1\nB. option2\nAnswer with the option letter only."Changes
Added new task variants:
mmmu_val_qwen(lmms_eval/tasks/mmmu/mmmu_val_qwen.yaml)pre_prompt: "Question: "post_prompt: "Answer with the option letter only."qwen3_vlfor proper prompt constructionmmstar_qwen(lmms_eval/tasks/mmstar/mmstar_qwen.yaml)Usage
# Use VLMEvalKit-compatible prompts for Qwen models python -m lmms_eval --model qwen2_5_vl \ --model_args pretrained=Qwen/Qwen2.5-VL-3B-Instruct \ --tasks mmmu_val_qwen,mmstar_qwenRelated Issues
Addresses #935, #932, #881, #901
Testing