Request for Official MVBench Evaluation Protocol for InternVideo2_5_Chat_8B

Hi OpenGVLab Team,

Thank you for releasing InternVideo2.5.

I am trying to reproduce the official MVBench score of 75.7 reported for OpenGVLab/InternVideo2_5_Chat_8B.
I am using the HuggingFace reference implementation as a starting point:

https://huggingface.co/OpenGVLab/InternVideo2_5_Chat_8B#%F0%9F%9A%80-how-to-use-the-model

To ensure my evaluation matches yours, I would appreciate clarification on the exact protocol, including:

• the number of frames and the frame-sampling strategy,
• preprocessing details (input resolution, dynamic tiling, thumbnail usage),
• the prompt template used for multiple-choice questions,
• decoding settings and the rule for extracting the predicted option.

Most importantly, could you share the official evaluation script or configuration used to obtain the reported 75.7 score?
If available, reference evaluation settings for other datasets (LongVideoBench, VideoMME, EgoSchema, Perception Test, LVBench) would also be very helpful.

Thank you in advance!

~Sergey

Start with the linked Hugging Face reference implementation for OpenGVLab/InternVideo2_5_Chat_8B and compare it with the reported MVBench score of 75.7. Done would be a documented official evaluation script or configuration covering frame sampling, preprocessing, prompts, decoding, and option extraction, with settings for the named datasets where available.