Details of architectural search

[Parsifal133]

I noticed that in your paper, you mentioned conducting a quick but reasonable architectural search using Video-LLaVA's training data. I'm interested in performing a similar search, but I noticed that the scale of Video-LLaVA's training data is quite large.

Could you please elaborate on the important details of the training, such as whether you used the entire dataset or a subset, and how many epochs you trained for?