Are all video-based checkpoints trained with 2 tokens?

[haodi19]

Hello, thank you for your great work. I noticed that in the open checkpoints, all checkpoints trained on video data have the compress type as 'mean' (or 'mean_concat', but I couldn't find the corresponding logic in the code). Are all video-based checkpoints, regardless of whether the training data is short or long videos, trained with 2 tokens?