TriviaQA is a reading comprehension dataset containing over 650K question-answer evidence triplets. TriviaQA includes 950K question-answer pairs from 662K documents from Wikipedia and the web. The dataset is more challenging than standard QA benchmarks such as the Stanford Question Answering Dataset (SQuAD) because the answers to the questions may not be directly available via span prediction and the context is long. The TriviaQA dataset consists of a subset of human-verified and machine-generated QA. Details can be found in the ACL 17 paper “TriviaQA: A Large-Scale Distant Supervision Challenge Dataset for Reading Comprehension"