MS MARCO Question Answering & Passage Ranking

## Overview MS MARCO (Microsoft MAchine Reading COmprehension) is a large-scale collection for deep learning in search, built from real anonymized Bing user queries. The QA configuration contains roughly 1M questions paired with human-generated answers and ~10 candidate passages each (with relevance labels). Format is Parquet (text modality), English only. Total size is in the 1M–10M row category across splits. ## Schema - query_id — int — unique identifier for the question - query — string — the natural-language question from Bing logs - query_type — string — category (e.g., DESCRIPTION, NUMERIC, ENTITY, LOCATION, PERSON) - passages — sequence — list of candidate passages, each with: passage_text (string), url (string), is_selected (int, relevance label) - answers — sequence[string] — human-written answer(s) to the query - wellFormedAnswers — sequence[string] — rewritten well-formed answers (subset of rows) ## Sources - HuggingFace: https://huggingface.co/datasets/microsoft/ms_marco — license: MIT (per Microsoft's MS MARCO terms of use / dataset card) - Original paper: Bajaj et al., "MS MARCO: A Human Generated MAchine Reading COmprehension Dataset" (arXiv:1611.09268) ## Methodology Questions were sampled from anonymized Bing search query logs and filtered to those expressing information need. For each query, ten passages were retrieved from the web index, and human judges wrote a natural-language answer based on those passages while also labeling which passage(s) they used (is_selected). Subsequent releases added well-formed answer rewrites and expanded the original 100K v1 set to ~1M questions in v2.1. ## Known gaps & limitations English-only; queries reflect Bing user distribution and time period (mid-2010s), so topical coverage skews to that era's web. Answers are extractive/abstractive from web passages and may contain factual errors inherited from sources. Not all rows have well-formed answers. The dataset has been extensively used for benchmark training and is known to overlap with many downstream IR evaluations — leakage risk is real. Microsoft's terms restrict use to non-commercial research in some interpretations; buyers should review the official MS MARCO terms at microsoft.github.io/msmarco for their use case. ## Intended use & out-of-scope - IS for: training and evaluating passage retrievers (BM25 baselines, dense retrievers like DPR/ColBERT), RAG pipelines, open-domain QA models, and reranker fine-tuning. - NOT for: training models that will be evaluated on BEIR, TREC-DL, or other MS MARCO–derived benchmarks without careful decontamination; not a source of ground-truth world knowledge (answers reflect 2016-era web). _Federated dataset: 12 parquet shards, 2.16 GB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: MS MARCO — Question Answering & Passage Ranking Microsoft's MS MARCO dataset: 1M+ real Bing questions with human-generated answers and passage relevance judgments. Foundational benchmark for retrieval, RAG, and QA systems.