chunked/nltk_chunking/model_doc_longformer/chunk_24.txt

All Longformer models employ the following logic for
global_attention_mask:

0: the token attends "locally",
1: the token attends "globally".