On the GLUE faq, they say:
Similarly, for SST, the data provided is already tokenized. We're working on obtaining a version that is not tokenized. Feel free to train on other distributions of SST, as long as they use the standard train/dev/test split.
Has anyone found an untokenized distribution?
The GLUE archive comes with an "original" folder with some untokenized data, but it definitely does not cover the entire tokenized dataset (it has 10,605 untokenized snippets which don't map to more than ~12k sentences, which is a lot less than the 67k tokenized training set that GLUE provides).