why use image input not video input in finetune?

[yaogang2060]

hi, I have two questions about build tokenized input to model:

1. Why using in build conversation, why not

2. Are you using defult Qwen3VL chat template? why not creating a chat template(maybe more convenient to describe mutli-view camera info?)?

Thank you!

[super-anova]

Hi, thanks for your question.
The use of video tokenizers is being investigated internally, so stay tuned for potential future releases!
As for the chat template, we indeed already use it. You can reference https://github.com/NVlabs/alpamayo/blob/main/src/alpamayo_r1/chat_template/conversation.py