The accuracy of llama2_2_3B on mvBench is a bit low.

[DuZzzs]

@xiaoqian-shen Hello, I ran the eval code, but the accuracy on the mvBench dataset is low. Is this normal? Thanks.
command：

num_gpus=1
torchrun --standalone --nnodes 1 --nproc_per_node $num_gpus eval/eval_mvbench.py --data_path ./data/opendata/MVBench/ \
--version llama3 --model_path checkpoints/LongVU_Llama3_2_3B

results:

Accuracy:  34.225
Task ccuracy {'Action Sequence': 25.0, 'Action Prediction': 27.5, 'Action Antonym': 62.5, 'Fine-grained Action': 22.5, 'Unexpected Action': 34.0, 'Object Existence': 52.0, 'Object Interaction': 29.5, 'Object Shuffle': 30.5, 'Moving Direction': 24.0, 'Action Localization': 23.5, 'Scene Transition': 28.5, 'Action Count': 41.5, 'Moving Count': 36.5, 'Moving Attribute': 42.5, 'State Change': 38.5, 'Fine-grained Pose': 26.5, 'Character Order': 41.5, 'Egocentric Navigation': 33.5, 'Episodic Reasoning': 30.5, 'Counterfactual Inference': 34.0}

[DuZzzs]

I can run eval_mvbench.py on a single 20G GPU, but it will report an out of memory error when running on 8 GPU cards with 80G.