lack of 13b pretrained model for videochat2 stage1,2,3

[pingpingpangpang]

i found that videochat1 provide pretrained model on both 7b and 13b, but for videochat2 i can see only 7b . please help to upload 13b pretrain model. many thanks.

[Andy1621]

Hi! Currently, we don't have a plan to release the 13B pretrained models, since the performances are similar in our experiments. And 13B requires more GPU sources for training. Recently, we plan to adapt VideoChat with other popular LLM, and we will upload the results if it's better.

[pingpingpangpang]

I succeed in depolying the videochat2 demo with vicuna 7b as you mentioned on the webpage. But it seems the performance is not as good as your web demo at https://vchat.opengvlab.com. I upload the snapshots for comparision. I am wondering if your web demo use the same model weights of vicuna 7b or any other better one?

[Andy1621]

Hi! We use the same old version of Vicuna-v0. I think the difference is caused by do_sample=False in the online demo. We do not conduct sampling for consistent response for online demo.

[pingpingpangpang]

I re-run the demo with code change "do_sample=False" in conversation.py, but the result is even worse , not like what you said.

[Andy1621]

Hi! Could you check whether you have loaded the stage3 checkpoint correctly? It runs normally in our environment, as shown in demo.ipynb.

[pingpingpangpang]

can we discuss this issue by some instant message tool, like wechat? give me your account. thanks.

[yinanhe]

You can add GV Assistant (her WeChat ID is gvxiaozhushou) or scan the WeChat QR code to join our group chat.