Ms Marco

MS MARCO (Microsoft MAchine Reading COmprehension) is a massive-scale dataset for Reading Comprehension and Passage Ranking, derived from real Bing search queries — containing 1M+ queries and partially human-generated answers, it is the standard benchmark for Neural Information Retrieval (IR).

Tasks

• Passage Ranking: Given a query, rank 1000 passages by relevance. (The "TREC" of the Deep Learning era).
• Answer Generation: Generate a natural language answer based on the retrieved passages.
• Key: Many queries have "No Answer" in the top passages.

Why It Matters

• Scale: Large enough to train data-hungry Transformers from scratch.
• Retrieval: The definitive benchmark for Dense Retrieval (DPR) and Re-ranking models (Cross-Encoders).
• Realism: Queries are short, noisy, and real ("how to cook pasta", "social security office hours").

MS MARCO is the search engine test — the definitive benchmark for teaching AI how to retrieve and rank relevant information from the web.