MS MARCO

Overview

MS MARCO (Microsoft MAchine Reading COmprehension) is a large-scale information retrieval benchmark built from real anonymized Bing search queries. Its passage ranking subset has become the default training and evaluation corpus for neural retrieval research — most published dense retrieval and sparse models report MS MARCO numbers, and many are trained on it directly.

Its scale is what makes it the standard stress test: the passage corpus contains roughly 8.8 million passages, large enough that memory footprint and index efficiency become real engineering constraints rather than afterthoughts.

Why It Recurs in Retrieval Work

  • Real queries. Drawn from Bing traffic, not synthesized — query distribution reflects actual user behavior, including the short, underspecified queries that dominate web search
  • Training scale. Hundreds of thousands of labeled query–passage pairs, enough to train neural rankers from scratch
  • Comparability. So widely used that a MS MARCO number is instantly comparable across papers
  • Benchmark for efficiency work. Corpus size makes it the natural testbed for Dimensionality Reduction, Vector Quantization, and Approximate Nearest Neighbor Search experiments — compression wins are measurable at this scale

Known Limitations

  • Sparse labels. Typically ~1 relevant passage marked per query; many genuinely relevant passages are unlabeled, so recall-oriented metrics understate true system quality
  • Binary relevance. The standard qrels are binary rather than graded, limiting NDCG’s usefulness compared to graded datasets like Amazon ESCI Dataset
  • Web-search domain. Models tuned on MS MARCO don’t automatically transfer to E-commerce Search, enterprise, or other verticals, motivating out-of-domain evaluation on separate benchmark suites

Comparison with Other Datasets

DatasetDomainScaleLabel type
MS MARCOWeb search (Bing)~8.8M passagesBinary, sparse
Amazon ESCI DatasetGeneral e-commerceVery large4-class (ESCI)
WANDS DatasetHome goods~42K pairs3-class
Home Depot Product Search RelevanceHome improvement~74K pairsContinuous 1–3
Natural QuestionsWikipedia QA~2.68M passagesBinary, sparse (~1.2/query)
TREC-COVIDBiomedical~171K docsGraded, deep (~493.5/query)

In This Vault

MS MARCO appears as the evaluation corpus across a broad slice of the retrieval notes — embedding compression, reranking, sparse retrieval, and metric definitions. Notably:

  • BEIR · MTEB — where MS MARCO-trained models get judged
  • LoTTE — built explicitly against MS MARCO’s head-heavy query distribution
  • BRIGHT · MIRACL · RTEB — the stress-test, multilingual, and anti-contamination benchmarks
  • Retrieval Benchmarks and Leaderboards — the landscape, and why MS MARCO’s ubiquity is now a contamination risk

Source