personachat-truecased

This dataset is a true-cased version of the PersonaChat dataset by Zhang et al. (2018). The original PersonaChat dataset is all lower case, and has extra space around each clause/sentence separating punctuation mark. This version of the dataset has more of a natural language look, with sentence capitalization, proper noun capitalization, and normalized whitespace. Also, each dialogue turn includes a pool of distractor candidate responses, which can be used by a multiple choice regularization loss during training.