LLaVA-Video精度测试对不齐

[linxid]

我测试了以下的模型的精度，模型精度和论文中宣称的不同。是我跑的有什么问题吗。
评测精度如下所示：

这是我评测脚本：

• LLaVA-Video-7B-Qwen2, fps32：

/opt/conda/envs/python3.10/bin/python -m accelerate.commands.launch --num_processes=8 \
    -m lmms_eval \
    --model llava_vid \
    --model_args pretrained=/tmp/pre-trained/lmms-lab/LLaVA-Video-7B-Qwen2,conv_template=qwen_1_5,max_frames_num=32,mm_spatial_pool_mode=average \
    --tasks videomme \
    --batch_size 1 \
    --log_samples \
    --log_samples_suffix llava_vid \
    --output_path ./logs/ 

• LLaVA-Video-7B-Qwen2, fps32：

/opt/conda/envs/python3.10/bin/python -m accelerate.commands.launch --num_processes=8 \
    -m lmms_eval \
    --model llava_vid \
    --model_args pretrained=/tmp/pre-trained/lmms-lab/LLaVA-Video-7B-Qwen2,conv_template=qwen_1_5,max_frames_num=32,mm_spatial_pool_mode=average \
    --tasks videomme \
    --batch_size 1 \
    --log_samples \
    --log_samples_suffix llava_vid \
    --output_path ./logs/ 

• LLaVA-Video-7B-Qwen2-Video-Only,fp32:

/opt/conda/envs/python3.10/bin/python -m accelerate.commands.launch --num_processes=8 \
    -m lmms_eval \
    --model llava_vid \
    --model_args pretrained=/tmp/pre-trained/lmms-lab/LLaVA-Video-7B-Qwen2-Video-Only,conv_template=qwen_1_5,max_frames_num=32,mm_spatial_pool_mode=average \
    --tasks videomme \
    --batch_size 1 \
    --log_samples \
    --log_samples_suffix llava_vid \
    --output_path ./logs/ 

• LLaVA-Video-7B-Qwen2-Video-Only,fp64:

/opt/conda/envs/python3.10/bin/python -m accelerate.commands.launch --num_processes=8 \
    -m lmms_eval \
    --model llava_vid \
    --model_args pretrained=/tmp/pre-trained/lmms-lab/LLaVA-Video-7B-Qwen2-Video-Only,conv_template=qwen_1_5,max_frames_num=64,mm_spatial_pool_mode=average \
    --tasks videomme \
    --batch_size 1 \
    --log_samples \
    --log_samples_suffix llava_vid \
    --output_path ./logs/