orgrctera/msmarco_document_ranking

MS MARCO Document Ranking

Dataset description and background

MS MARCO (MicroSoft MAchine Reading COmprehension) is a large-scale collection originally introduced for machine reading comprehension and question answering. Over time it has become a standard benchmark for information retrieval under abundant training data: hundreds of thousands of queries with human relevance signals, aligned with real web search behavior.

The document ranking track uses a document-level corpus derived from the same ecosystem as MS MARCO passage ranking. The official resource describes a corpus on the order of ~3.2M documents, with training queries on the order of hundreds of thousands, development and leaderboard test query sets, and TREC-style qrels (query–relevance judgments) for training and development. For training, passage-level labels are mapped to document IDs under the assumption that a document containing a judged-relevant passage is treated as a relevant document—supporting transfer between passage-focused and document-focused retrieval research.

This Hugging Face dataset is a CTERA-packaged view of that task: each row pairs a natural-language query with structured labels/metadata suitable for retrieval and RAG benchmarking (see Data fields below).

Official MS MARCO ranking resources (corpus, qrels, leaderboards): MS MARCO — Datasets for Document and Passage Ranking

Task: document ranking / retrieval

This dataset supports ad-hoc document retrieval / ranking: given a query, a system should rank documents from a collection by relevance. In research settings this is often split into:

• Full ranking (retrieval): score or retrieve from the full document collection (official submissions allow a bounded number of documents per query, e.g. up to 100 in the MS MARCO document ranking setup).
• Top‑k reranking: rerank a fixed candidate list (e.g. top‑100 candidates from a first-stage retriever)—a common production pattern of “retrieve, then rerank.”

Evaluation in the MS MARCO ranking leaderboards is typically reported with MRR@10 (Mean Reciprocal Rank at rank 10) for the document ranking task, alongside standard TREC-style analyses where applicable. (Consult the official leaderboard and TREC Deep Learning track materials for the exact metric definitions used in a given campaign.)

Data fields

Parquet splits expose three columns:

Column	Description
input	The query text (string).
expected_output	JSON string with relevance information; format differs by split (see examples).
metadata	JSON string with identifiers and benchmark tags (benchmark_name, split, query_id, etc.).

Splits: train, dev, and test are provided. The test split may use held-out or empty labels in expected_output for blind evaluation workflows—check the sample below.

Examples

Train (expected_output includes query id and judged relevant document id(s) as JSON):

{
  "input": ")what was the immediate impact of the success of the manhattan project?",
  "expected_output": "{\"qid\": \"1185869\", \"qrels\": [\"D59219\"]}",
  "metadata": "{\"query_id\": \"1185869\", \"split\": \"train\", \"benchmark_name\": \"msmarco_document_ranking\", \"benchmark_type\": \"base_rag\", \"sub_benchmark\": \"document_ranking\"}"
}

Dev (expected_output is a JSON-encoded list of relevant document ids):

{
  "input": "does xpress bet charge to deposit money in your account",
  "expected_output": "[\"D1987644\"]",
  "metadata": "{\"query_id\": \"174249\", \"split\": \"dev\", \"benchmark_name\": \"msmarco_document_ranking\", \"benchmark_type\": \"base_rag\", \"sub_benchmark\": \"document_ranking\"}"
}

Test (labels may be withheld; example shows an empty list):

{
  "input": "how to display how.close you are to.cell.tower",
  "expected_output": "[]",
  "metadata": "{\"query_id\": \"355339\", \"split\": \"test\", \"benchmark_name\": \"msmarco_document_ranking\", \"benchmark_type\": \"base_rag\", \"sub_benchmark\": \"document_ranking\"}"
}

References

Foundational MS MARCO paper (cite when using MS MARCO-derived data)

Payal Bajaj, Daniel Campos, Nick Craswell, Li Deng, Jianfeng Gao, Xiaodong Liu, Rangan Majumder, Andrew McNamara, Bhaskar Mitra, Tri Nguyen, Mir Rosenberg, Xia Song, Alina Stoica, Saurabh Tiwary, Tong Wang. MS MARCO: A Human Generated MAchine Reading COmprehension Dataset. arXiv:1611.09268, 2016.

Abstract (short): MS MARCO introduces a large-scale machine reading comprehension dataset built from real Bing queries, with human-generated answers and millions of passages from retrieved web documents. The work defines multiple tasks of varying difficulty, including passage ranking—establishing MS MARCO as a benchmark for realistic, large-scale QA and IR research.

• Paper: arXiv:1611.09268

@article{bajaj2016ms,
  title={Ms marco: A human generated machine reading comprehension dataset},
  author={Bajaj, Payal and Campos, Daniel and Craswell, Nick and Deng, Li and Gao, Jianfeng and Liu, Xiaodong and Majumder, Rangan and McNamara, Andrew and Mitra, Bhaskar and Nguyen, Tri and others},
  journal={arXiv preprint arXiv:1611.09268},
  year={2016}
}

Official dataset and leaderboard documentation

• MS MARCO — Datasets for Document and Passage Ranking Leaderboards — corpus files, qrels, submission formats, and task description.
• MS MARCO — Submission / evaluation — ranking submission conventions.
• TREC Deep Learning Track — blind evaluation and community benchmarks related to MS MARCO ranking tasks.

Related code and corpora (Microsoft)

• microsoft/MSMARCO-Document-Ranking — pointers and tooling around the document ranking collection.

Terms and licensing

The underlying MS MARCO data is subject to Microsoft’s terms for non-commercial research as published on the official MS MARCO site; review the Terms and Conditions on the official datasets page before use in products or redistributions.

This Hugging Face dataset card describes the CTERA-formatted release; when publishing results, cite MS MARCO appropriately and follow the original license / usage constraints for the source corpus and judgments.

Acknowledgments

Dataset packaging for this repository is maintained by CTERA. MS MARCO is provided by Microsoft and the broader IR community; see the official site for credits and contact information.