interleaved understanding, generation video/web data during training

Thanks for sharing your brilliant work! I have two question here:
1. Could you please clarify the data format in interleaved understanding and interleaved generation (video/web) during training? Are there any concrete data example available?
2. For interleaved training data, how to compute loss for certain part? for example, the training sample is (image1, text1, image2, text2, image3), I want to compute loss on predicting text2 and image3, how should the code be adapted?

The issue does not name files, tests, or entry points. First clarify the interleaved video/web training data format and the intended loss mask for the (image1, text1, image2, text2, image3) example. Done means documented concrete data examples and an agreed implementation path for computing loss only on text2 and image3.

Written by the indexing model from the issue text.