About training epochs and time for each stage

[raymond-huang-sony]

Thank you for the brilliant work!

I would like to run the training of VILA-U for a better understanding but I failed to find the detailed training setup in the paper. It seems like the default number of epochs set in the pre-training script is one. Could you confirm whether this is the same as for the models reported in the paper?

Besides, it would be great if you could suggest the training time for each stage, i.e. training the vision tower, pre-training, and supervised fine-tuning. P.S. I found the overall training cost reported in #2 .

Thank you.