[GH-ISSUE #117] Request for Official MMMU-val Evalution Script #96

Open
opened 2026-10-26 21:11:52 +00:00 by bsakaguchi · 0 comments
Owner

Originally created by @willburwwb on GitHub (Mar 18, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-VL2/issues/117

Hello, I evaluated the inference accuracy of the DeepSeek-VL2-smal model using the MMMU (https://github.com/MMMU-Benchmark/MMMU) official benchmark script. However, I noticed that the results differ significantly from the accuracy reported in the paper.

Details are as follows:

  • Evalution dataset: MMMU val
  • Achieved accuracy : 35% ~ 37%
  • Accuracy reported in the paper: 48%

To better reproduce the results in the paper, could you please provide the official evalution script for the MMMU-val dataset or clarify if there are any special configurations or preprocessing steps required during evalution?

Originally created by @willburwwb on GitHub (Mar 18, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-VL2/issues/117 Hello, I evaluated the inference accuracy of the DeepSeek-VL2-smal model using the MMMU (https://github.com/MMMU-Benchmark/MMMU) official benchmark script. However, I noticed that the results differ significantly from the accuracy reported in the paper. Details are as follows: - Evalution dataset: MMMU val - Achieved accuracy : 35% ~ 37% - Accuracy reported in the paper: 48% To better reproduce the results in the paper, could you please provide the official evalution script for the MMMU-val dataset or clarify if there are any special configurations or preprocessing steps required during evalution?
Sign in to join this conversation.
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
deepseek-ai/DeepSeek-VL2#96
No description provided.