Glossary: NLP with Transformers
-100 label
Also written: -100, ignore index
A label value that PyTorch's cross-entropy loss ignores. In NER it goes on special tokens and on every word piece after the first, so only each word's first piece is trained.