PoetryMTEB/Poetry50KRetrieval

Poetry 50K Retrieval

BEIR-style English poetry retrieval benchmark for PoetryMTEB / MTEB, packaged from John Foley's dissertation datasets (Poetry: Identification, Entity Recognition, and Retrieval, UMass Amherst, 2019).

Task: given a short topical query (e.g. love, satire, haiku), rank OCR poetry pages from the Poetry 50K collection by graded relevance.

Item	Description
Dataset version	1.0.0
Hub repo	PoetryMTEB/Poetry50KRetrieval
Task	Retrieval (graded, score ∈ {0,1,2})
Language	English (en), historical OCR
Splits	test only (MTEB Retrieval)
Corpus	Poetry 50K deduplicated unique pages
Queries	26 topical keywords
Judgments	Crowdsourced MTurk + adjudication (mturk.max / mturk.min)
Metrics	nDCG@10, MAP, Recall@k
Upstream	https://ciir.cs.umass.edu/downloads/poetry/

Counts

Config	Split	#rows
corpus	test	570,930
queries	test	26
qrels_max	test	1,170
qrels_min	test	1,170

• Judged documents (union of max/min): 1,326
• Avg. judged docs / query (max): 45.0
• Corpus coverage of judged ids: 1,149 / 1,326

Score histograms (after mapping -1 → 0)

qrels_max (optimistic): {0: 207, 1: 353, 2: 787}
qrels_min (pessimistic): {0: 585, 1: 508, 2: 254}
Raw Foley labels before mapping: max={-1: 124, 0: 83, 1: 353, 2: 787}, min={-1: 247, 0: 338, 1: 508, 2: 254}

Configs

corpus

Deduplicated Poetry 50K pages (poetry50k.dedup.jsonl.gz, MD5 be61efeff6ae2b1ffb86519aeeabf427).

Field	Type	Description
id	string	Document id {book}/{page} (Internet Archive book id + page)
text	string	Retrieval document: OCR page text (\t→space, lines preserved)
book	string	Internet Archive book identifier
page	int64	Page number within the book
score	float64	Upstream poetry-detection RF score
n_duplicates	int64	Size of duplicate cluster
duplicates	string	JSON list of duplicate book/page ids

queries

Field	Type	Description
id	string	Topic id (underscore form), e.g. mothers_day
text	string	Retrieval query: humanized topic (mothers day)

Topics (26): ancestors, buddhism, dance, death, doubt, easter, education, ethnicity, flowers, friendship, funny, graduation, haiku, inspiration, jabberwocky, love, mothers_day, photography, reading, relationships, religion, satire, sorrow, teacher, war, wedding

Upstream releases no separate query file; topics are the TREC qrel query ids. We materialize queries.text by replacing _ with spaces.

qrels_max (default)

Optimistic aggregation: if any annotator marked relevant, keep the higher/optimistic label (mturk.max.qrel).

qrels_min

Pessimistic aggregation: if any annotator marked non-relevant, reflect that (mturk.min.qrel).

Field	Type	Description
query-id	string	Matches queries.id
corpus-id	string	Matches corpus.id
score	float64	Graded relevance 0 / 1 / 2

Relevance scale

Score	Meaning (Foley MTurk)
2	Highly relevant
1	Relevant / partially relevant
0	Not relevant or Foley raw -1 (adjudicated non-poetry / unusable), mapped to 0 for MTEB

Default Hub / evaluation config: qrels_max.

How to load

from datasets import load_dataset

corpus = load_dataset("PoetryMTEB/Poetry50KRetrieval", "corpus", split="test")
queries = load_dataset("PoetryMTEB/Poetry50KRetrieval", "queries", split="test")
qrels = load_dataset("PoetryMTEB/Poetry50KRetrieval", "qrels_max", split="test")   # default
qrels_min = load_dataset("PoetryMTEB/Poetry50KRetrieval", "qrels_min", split="test")

print(len(corpus), queries[0], qrels[0])

# Build a simple qrels dict for evaluation
from collections import defaultdict
qrel_dict = defaultdict(dict)
for r in qrels:
    qrel_dict[r["query-id"]][r["corpus-id"]] = r["score"]

Construction (summary)

1. Download Poetry 50K dedup JSONL + mturk.max.qrel / mturk.min.qrel from CIIR.
2. Corpus id = book/page; normalize OCR text for retrieval.
3. Queries = unique qrel topic ids; text = topic with _ → space.
4. Map Foley score -1 → 0; keep 0/1/2 graded labels.
5. Package BEIR configs for PoetryMTEB (test-only).

License / attribution

Upstream page does not attach a machine-readable SPDX license; this packaging uses license: other and requires citing Foley's dissertation. Internet Archive book scans remain subject to their original access terms. Redistribution of OCR text is for research / evaluation under PoetryMTEB.

Citation

@phdthesis{foley2019poetry,
  title  = {{Poetry: Identification, Entity Recognition, and Retrieval}},
  school = {{University of Massachusetts Amherst}},
  author = {{Foley, John}},
  year   = {2019},
  url    = {{https://ciir.cs.umass.edu/downloads/poetry/}}
}

@misc{poetry50k_retrieval_poetrymteb,
  title        = {{Poetry 50K Retrieval: Topical English Poetry Page Retrieval}},
  author       = {{PoetryMTEB Contributors}},
  year         = {2026},
  version      = {1.0.0},
  publisher    = {{Hugging Face}},
  url          = {{https://huggingface.co/datasets/PoetryMTEB/Poetry50KRetrieval}},
  note         = {{PoetryMTEB BEIR-style packaging of Foley Poetry 50K + MTurk qrels}}
}