Data-to-text generation models are used to generate a textual description of structured data (e.g. a table), for example, to create basketball game summaries given box score statistics. Data-to-text tasks can also help evaluate the extent to which language models may “hallucinate” – generating text that is fluent but not faithful to the source. To support this R&D, Google Research has released the ToTTo Dataset, an open-domain, English table-to-text dataset with more than 120,000 training examples. ToTTo is built around the task of producing a one-sentence description given a Wikipedia table and set of highlighted cells.