Shroedingercat/doubting-ms-marco

Doubting MS MARCO (v0.1-preview)

A cleaner, more densely annotated MS MARCO resource for training and evaluating passage retrievers. It provides over 8 million training-positive pairs and 238,184 strict-positive pairs covering 6,850 development queries, keyed by the original MS MARCO query and passage IDs. A frozen-LLM search agent combines lexical and dense retrieval with adaptive query reformulation to discover additional answer passages. Separate calls to the same model assess each proposed passage and every official positive.

In a 400-query human audit, 91% of our randomly sampled strict positives fully answered the query, compared with 68.8% of official positives. Across four evaluated retrievers, training on our positives consistently improves zero-shot transfer to BEIR and MS MARCO retrieval quality compared with training on official positives. The annotations accompany the paper Doubting MS MARCO: Cleaner Labels for Retrieval Training and Evaluation.

The package holds ids and annotations only. It contains no query or passage text; join the ids with your own copy of MS MARCO (see below). This is a preview release: its files are frozen by the checksums in SHA256SUMS, and a later release that changes any content will carry a new version.

Contents

Config	Split	Rows	One row is
qrels (default)	train	8,100,490	a training positive: an agent-proposed passage with grade 2 or 3 and a finite WOOD score
qrels	dev	238,184	a strict development positive: grade 3
judgments	train	10,394,573	a graded pair: 9,864,256 agent proposals and 530,317 official positives
judgments	dev	1,423,357	a graded pair: 1,422,376 agent proposals and 981 official positives the agent did not propose
train_queries	train	500,607	a query of the training collection, with counts and its selected training positive
dev_queries	dev	6,980	a development query, with counts and its strict-subset flag
generated_answers	train	499,327	the model-generated answer used to compute WOOD for that query

Companion files: trec/dev.strict.qrels (the strict development qrels in TREC format), ids/ (the training collection, the excluded training ids with the reason, all development ids and the strict development subset), examples/, MANIFEST.json (version, row counts, source and code checksums) and SHA256SUMS.

Coverage

Train. The collection has 500,607 MS MARCO training queries: the 502,910 queries assigned for annotation minus 24 whose text equals that of a held-out development or test query and 2,279 whose text repeats an earlier query's, with texts compared in lower case and with punctuation removed (ids/train_excluded.tsv).

• The agent proposed passages for 499,327 queries: 9,864,256 graded query-passage pairs. 1,280 queries have no agent row. They stay in train_queries with agent_row = false and zero agent counts; their official positives are graded and one of them is the selected fallback positive.
• Grades of the agent's proposals: 0: 1,068,976, 1: 694,761, 2: 2,286,874, 3: 5,813,645. 36 proposals have no finite WOOD score (wood is null).
• Training positives (grade 2 or 3 and finite WOOD): 8,100,490 pairs on 496,190 queries, a mean of 16.18 and a median of 15 per collection query, with every collection query in the denominator. 3,137 queries have agent proposals but no qualifying pair. Grade 3 with finite WOOD: 5,813,620 pairs on 481,844 queries. Grade 3 without the WOOD condition: 5,813,645 pairs.
• Official MS MARCO positives, graded separately: 530,317 pairs (grades 0: 34,505, 1: 17,554, 2: 77,943, 3: 400,315). When the agent also proposed an official positive, both rows are kept, with their own origin and grade.
• Selected training positive, one per query: 496,190 chosen by WOOD and 4,417 official fallbacks for the queries without a qualifying pair. A fallback is marked positive_source = official_fallback, keeps whatever grades and WOOD score it has (often none), and is not counted among the training positives above.

Dev. All 6,980 queries of the MS MARCO dev set ("dev small"). 1,423,357 graded pairs: 1,422,376 agent proposals and 981 official positives the agent did not propose; the dev set has 7,437 official positive pairs in total. The strict qrels hold 238,184 grade-3 pairs on 6,850 queries (mean 34.12 and median 14 per query over all 6,980). The other 130 queries have no grade-3 pair: 124 have graded proposals none of which received grade 3, and 6 have no agent proposal. They remain in dev_queries (strict_subset = false) and in ids/dev_all.txt; a missing grade 3 does not show that no answering passage exists.

How the annotations were made

• Model. A frozen Qwen3.6-35B-A3B served the search agent, the grading calls and WOOD.
• Search. For each query the agent searched lexical and dense indexes of the passage collection, inspected results, reformulated the query and proposed passages. agent_rank is the passage's position in the proposal list. The development run searched more broadly: its proposal lists reach 400 passages per query, against 60 on train.
• Grades. A separate call without the search history graded each pair: 0 irrelevant, 1 marginally related, 2 relevant, 3 directly answering (an answer supported by the passage itself, including paraphrase or a short unambiguous inference). Training grades see only the query and the passage (grading = query_only). Development grades also see the MS MARCO reference answer as a fallible hint and accept other correct answers (grading = answer_hinted). Official positives were graded by the same rule as the agent's proposals of the same split. Train and dev grades therefore come from different settings and are not directly comparable.
• WOOD (Weighted Output Odds Differential; training only). The model writes a short answer from up to three proposed passages. WOOD of a passage is the mean gain in log-likelihood of that answer's tokens when the passage is given, compared with the query alone. The generated answers are in generated_answers; they are model output derived from passages, not MS MARCO reference answers.
• Selected positive. The highest finite WOOD among the query's grade 2-3 proposals; when no proposal qualifies, a deterministic official positive.
• Development set. 300 development queries come from an earlier pilot run of the same search and grading recipe (acquisition_batch = dev-pilot-2026-09); the others were run in October 2026 (dev-2026-10).
• Acquisition batches (train). cluster-bf16-2026-09: 99,358 queries, served in bf16 on a local cluster in September 2026. cluster-bf16-2026-10: 384,933 queries, the same recipe in October 2026. api-fp8-2026-10: 16,316 queries whose search and grades ran on an fp8 API deployment of the same model (WOOD on the cluster in bf16).

Schema

All ids are strings exactly as in MS MARCO. Grades are integers 0-3. Null means "not available", never zero.

Config	Column	Type	Meaning
qrels	query_id, passage_id	string	the pair
	grade	int8	2 or 3 on train, 3 on dev
judgments	query_id, passage_id	string	the pair; unique together with origin
	origin	string	agent (proposed by the agent) or official (official MS MARCO positive graded on its own)
	grade	int8	0-3
	grading	string	query_only (train) or answer_hinted (dev)
	agent_rank	int32	position in the agent's proposal list; null for official rows
	wood	float64	WOOD score; null for official rows, for dev, and for the 36 train proposals without a finite score
	train_eligible	bool	train: grade 2 or 3 and finite WOOD (the qrels train rule); null on dev and for official rows
	official_positive	bool	the pair is a positive in the original MS MARCO qrels of that split
train_queries	query_id	string	collection query
	acquisition_batch	string	see above
	agent_row	bool	the agent proposed at least one passage
	n_agent, n_eligible, n_eligible_grade3, n_official	int32	proposals; training positives; those with grade 3; official positives
	positive_id, positive_source	string	selected training positive; wood or official_fallback
	positive_agent_grade, positive_official_grade, positive_wood	int8, int8, float64	its grade as an agent proposal, its grade as an official positive and its WOOD; null where it is not one
dev_queries	query_id, acquisition_batch	string	development query; dev-pilot-2026-09 or dev-2026-10
	strict_subset	bool	has at least one grade-3 pair
	n_agent, n_official, n_judged, n_strict	int32	agent proposals; official positives; graded pairs; grade-3 pairs
generated_answers	query_id, answer	string	the generated answer used for WOOD

Loading

from datasets import load_dataset

repo = "Shroedingercat/doubting-ms-marco"  # a local copy of the folder works as well
qrels = load_dataset(repo, "qrels")                              # DatasetDict with train and dev
dev_judgments = load_dataset(repo, "judgments", split="dev")
train_queries = load_dataset(repo, "train_queries", split="train")
selected = train_queries.filter(lambda r: r["positive_source"] == "wood")

examples/load.py prints a few rows of every config. Large splits can be streamed with streaming=True.

Joining the original text

The ids refer to the MS MARCO passage ranking v1 collection (8,841,823 passages; the ids of collection.tsv, of ir_datasets msmarco-passage and of BEIR msmarco), its training queries and its 6,980-query "dev small" set. Obtain the text from the original source under its terms, for example:

import ir_datasets
docs = ir_datasets.load("msmarco-passage").docs_store()
queries = {q.query_id: q.text for q in ir_datasets.load("msmarco-passage/dev/small").queries_iter()}
row = dev_judgments[0]
print(queries[row["query_id"]], "|", docs.get(row["passage_id"]).text)

examples/join_text.py does the same from the official TSV files (collection.tsv, queries.train.tsv, queries.dev.small.tsv) and needs only pyarrow; for the judgments config it keeps the origin column, because one pair can carry both an agent grade and an official-positive grade.

Evaluating on the strict dev qrels

python examples/evaluate.py --qrels trec/dev.strict.qrels --run my_run.trec
python examples/evaluate.py --export-train-qrels train.qrels --package .   # TREC qrels from the qrels train split

evaluate.py reports MRR@10, nDCG@10 and Recall@100 over the 6,850 queries of the strict subset, the usual convention for queries without a relevant pair; ids/dev_all.txt lists all 6,980. Unjudged passages count as non-relevant, as in standard practice; they are not verified negatives.

Citation

Paper: Doubting MS MARCO: Cleaner Labels for Retrieval Training and Evaluation. The citation entry will be added with the paper.