Exploring Token Probabilities as a Means to Filter GPT-3’s Answers

To build better GPT-3-powered chatbots

Log probabilities produced by GPT-3 for each token that makes up the displayed sentence, made up to test the system. The picture was composed by the author from a screenshot of a web app written to carry out these tests, linked near the end of the article.

As powerful language models become increasingly prevalent, the need for control over the content they generate becomes ever more pressing. These models, trained on massive amounts of text data, have the ability to generate highly convincing written content, from news articles to social…