Question regarding InternVideo2+VideoChat2 results on MVBench

[jpan72]

Hello authors,

Table 16 in the InternVideo2 paper reports a score of 60.9 on MVBench, with VideoChat2 + InternVideo2s3-1B + Mistral-7B.

However, the highest score on MVBench leaderboard is 60.4 (VideoChat2_mistral) as of today.

I am just curious -

1. Why is the 60.9 result not reported to the MVBench leaderboard yet?
2. Are these two results referring to the same combination of models? If not, what's the difference?

Thanks!

[jpan72]

For question 2, is it because

UMT-L + VideoChat2 + Mistral-7B=60.4 --> Result reported as MVBench top-1 InternVideo2 + VideoChat2 + Mistral-7B = 60.9 --> Result reported in InternVideo2 paper

Another question: 3. Could you please share model weights of InternVideo2 during VideoChat2 training? Something like this table https://github.com/OpenGVLab/Ask-Anything/blob/main/video_chat2/README.md#model but the InternVideo2 version instead of the UMT version?

Appreciate it!

[bbkaeul]

Have you got the model weights of InternVideo2 during VideoChat2 training?

[yinanhe]

Hi, the weight of InternVideo2 stage3 can be found here and here