TriviaQA is a large-scale reading-comprehension / question-answering dataset introduced by Joshi et al. (ACL 2017). It contains over 650K question-answer-evidence triples (about 95K question-answer pairs authored by trivia enthusiasts) with independently gathered evidence documents (about six evidence documents per question on average). The dataset provides both a reading-comprehension (RC) version (contexts where answers appear) and an unfiltered / open-domain style version (where not all retrieved documents necessarily contain the answer). TriviaQA was designed to be more challenging than prior RC datasets: questions are often compositional, exhibit high syntactic/lexical variability relative to answer-evidence sentences, and frequently require cross-sentence reasoning. The original paper provides RC and open-domain splits and baselines; data and downloads are available from the project page and via Hugging Face.
No results indexed yet — be the first to submit a score.
Submit a checkpoint and a reproduction script. We will run it, publish the score, and — if it takes the top — annotate the step on the progress chart with your name.