We introduce a large scale MAchine Reading COmprehension dataset, which we name MS MARCO. The dataset comprises of 1,010,916 anonymized questions---sampled from Bing's search query logs---each with human generated answer and 182,669 completely rewritten answers. In addition, the contains 8,841,823 passages---extracted 3,563,535 web documents retrieved by Bing---that provide information necessary for curating natural language A question in MARCO may have multiple answers or no at all. Using this propose three different tasks varying levels difficulty: (i) predict if is answerable given set context passages, extract synthesize as would (ii) generate well-formed (if possible) based on passages that can be understood passage context, finally (iii) rank question. size fact questions are derived real user queries distinguishes other well-known publicly available datasets machine reading comprehension question-answering. believe real-world nature makes it attractive benchmarking question-answering models.

MS MARCO: A Human Generated MAchine Reading COmprehension Dataset