Request for Inference Parameters on VideoLLava

[adrianwestmoon]

Thank you for the fantastic work on VideoLLava!

I have been attempting to replicate the results presented in the benchmark, specifically on VideoMME, but I am consistently observing a performance drop of about 5 points compared to the reported results by using the pretrained model from LanguageBind/Video-LLaVA-7B-hf.

Could you kindly provide the inference parameters or any specific settings you used for the evaluation on VideoMME? This would greatly help me in understanding and hopefully resolving the discrepancy.

Thank you in advance for your assistance!