This paper addresses the issue of chit-chat models lacking specificity, consistent personality, and captivation. The authors propose a method to enhance chit-chat by conditioning on profile information, which is collected through a crowdsourcing dataset called PERSONA-CHAT. This dataset consists of 162,064 utterances between crowdworkers who were randomly paired and asked to act as given personas. The models are trained to condition on their own or their partner's profile, improving the quality of the dialogues as measured by next utterance prediction. The paper also introduces a new evaluation metric, profile prediction, which shows that models can be trained to predict the profiles of interlocutors based on their dialogues. The results demonstrate that models conditioned on personas are more engaging and consistent, and that the PERSONA-CHAT dataset is more effective than other datasets like OpenSubtitles and Twitter in training chit-chat models.This paper addresses the issue of chit-chat models lacking specificity, consistent personality, and captivation. The authors propose a method to enhance chit-chat by conditioning on profile information, which is collected through a crowdsourcing dataset called PERSONA-CHAT. This dataset consists of 162,064 utterances between crowdworkers who were randomly paired and asked to act as given personas. The models are trained to condition on their own or their partner's profile, improving the quality of the dialogues as measured by next utterance prediction. The paper also introduces a new evaluation metric, profile prediction, which shows that models can be trained to predict the profiles of interlocutors based on their dialogues. The results demonstrate that models conditioned on personas are more engaging and consistent, and that the PERSONA-CHAT dataset is more effective than other datasets like OpenSubtitles and Twitter in training chit-chat models.