The evaluation results of Qwen2.5-VL-7B on some benchmarks are inconsistent with that on the leaderboard

Hi, we attempted to reproduce the evaluation of Qwen2.5-VL-7B, but the results were inconsistent with those reported on the leaderboard.

For example, the evaluation score for MathVerse_MINIVisionOnly is 31.2, while the leaderboard is 41.1. The evaluation score of LogicVista is 40.5, while the leaderboard is 47.9. Here is the detailed information:

# MathVerse_MINIVisionOnly
[2025-03-28 17:57:00] INFO - run.py: main - 414: 
--------------  ------------------
split           Vision Only
Overall         31.218274111675125
Solid Geometry  10.084033613445378
Plane Geometry  33.52941176470588
Functions       39.62264150943396
Expression      15.625
Analytic        34.883720930232556
Applied         29.357798165137616
Angle           33.67875647668394
Area            20.87912087912088
Property        40.845070422535215
Volume          9.803921568627452
Coordinate      62.5
Length          36.26373626373626
--------------  ------------------

# LogicVista
[2025-03-28 17:59:30] INFO - run.py: main - 414: 
-  ----------  ---  ---  -------
0  Overall     447  181  40.4922
1  inductive   107   38  35.514
2  deductive    93   51  54.8387
3  numerical    95   33  34.7368
4  spatial      78   26  33.3333
5  mechanical   74   33  44.5946
-  ----------  ---  ---  -------

However, some results are consistant with the results on the leaderboard, for example, the evaluation score of WeMath is 35.6 (36.2 on leaderboard).

-------------------------------------------  ------------
Model                                        nan
Score (Strict)                               35.62%
InsufficientKnowledge (Strict)               37.33% (196)
InadequateGeneralization (Strict)            14.86% (78)
CompleteMastery (Strict)                     28.19% (148)
RoteMemorization (Strict)                    41.04% (103)
Score (Loose)                                52.57%
InsufficientKnowledge (Loose)                37.33% (196)
InadequateGeneralization (Loose)             14.86% (78)
CompleteMastery (Loose)                      45.14% (237)
RoteMemorization (Loose)                     5.58% (14)
One-step(S1)                                 68.72%
Two-step(S2)                                 50.56%
Three-step(S3)                               41.82%
Understanding and Conversion of Units        78.37%
Angles and Length                            55.79%
Calculation of Plane Figures                 73.53%
Understanding of Plane Figures               62.65%
Calculation of Solid Figures                 72.70%
Understanding of Solid Figures               64.28%
Basic Transformations of Figures             47.46%
Cutting and Combining of Figures             70.90%
Direction                                    79.05%
Position                                     73.01%
Route Map                                    58.79%
Correspondence of Coordinates and Positions  70.00%
-------------------------------------------  ------------

The results are obtained by running torchrun --nproc-per-node=8 --master_port 29500 run.py --data LogicVista --model Qwen2.5-VL-7B --mode all --api-nproc 8

Could you please clarify what environment was used for the evaluation of the Qwen2.5-VL-7B model on the leaderboard? Additionally, would it be possible to disclose the detailed evaluation results?