The paper introduces MS MARCO, a large-scale machine reading comprehension (MRC) dataset. The dataset comprises 1,010,916 anonymized questions, each with a human-generated answer, and 182,669 completely human-rewritten answers. It also includes 8,841,823 passages extracted from 3,563,555 web documents. The questions are derived from Bing's search query logs, making them more representative of real-world information needs. The dataset is designed to address the limitations of existing MRC datasets, such as the lack of real-world complexity and the need for models to handle noisy and conflicting information. The paper proposes three tasks: predicting if a question is answerable, generating a well-formed answer, and ranking retrieved passages. The authors present benchmarking results using various models and discuss the challenges and future directions for the dataset.The paper introduces MS MARCO, a large-scale machine reading comprehension (MRC) dataset. The dataset comprises 1,010,916 anonymized questions, each with a human-generated answer, and 182,669 completely human-rewritten answers. It also includes 8,841,823 passages extracted from 3,563,555 web documents. The questions are derived from Bing's search query logs, making them more representative of real-world information needs. The dataset is designed to address the limitations of existing MRC datasets, such as the lack of real-world complexity and the need for models to handle noisy and conflicting information. The paper proposes three tasks: predicting if a question is answerable, generating a well-formed answer, and ranking retrieved passages. The authors present benchmarking results using various models and discuss the challenges and future directions for the dataset.