How to generate CSV to train on a new dataset

[mario-m-g]

Hello,

I am trying to train TalkNet on a different dataset but I don't understand how both CSV train_orig and train_loader work.
I have already seen #30 and #10. But I still don't get what some of the fields mean.

1. What do entity_id, label_id and instance_id mean?
2. A line in train_loader is, for instance: 8aMv-ZGD4ic_1500_1560:21 87 25.0. I guess 25 is fps an 87 is the number of frames in that clip. And 21, what does that represent?
3. How is train_loader generated from train_orig? Is there any code for that? Is it created somewhere in the provided code and I have not seen it?

Thanks in advance!