About Training Settings

[baifanxxx]

Hi,

Thank you for your good work for the community. Can I ask about the settings for training VILA-U? For example, GPU type, quantity, and days.

Best regards,
BAI Fan

[zhuoyang20]

Hi @baifanxxx,

Thank you for your interest!

We trained VILA-U on 16 nodes of A100 for around 20K GPU hours. This includes the training of the unified vision tower, multi-modal pre-training and sft. Please feel free to reach out if you have additional questions.

Best,
Zhuoyang

[John-Ge]

I am not sure if the details of sft data are in the paper. I do not find it. Could this be released? Thank you!