chunked/nltk_chunking/_perplexity/chunk_26.txt

We don't want the log-likelihood for the tokens we're just treating
as context to be included in our loss, so we can set these targets to -100 so that they are ignored.