What are separator tokens and how are they used in training?

Separator tokens are used to distinguish different time-windows and examples during training. They help the model understand the context and boundaries between different time-series segments, particularly when using packed examples.