PoetryMTEB/ClassicalPoetryRetrieval

Classical Poetry Retrieval

BEIR-style multi-aspect classical Chinese poetry retrieval for PoetryMTEB / MTEB.

Chinese queries retrieve classical poems along four aspects (emotion / intent / theme / thought), with graded relevance (score ∈ {0,1,2,3}).

Item	Description
Dataset version	1.2.0
Hub repo	PoetryMTEB/ClassicalPoetryRetrieval
Task	Retrieval (graded, score ∈ {0,1,2,3})
Language	Classical Chinese / Chinese (zh)
Aspects	emotion · intent · theme · thought
Splits	train (LLM), silver (LLM test), gold (human RAA)
Metrics	nDCG@10, MAP, Recall@k

Counts

qrels rows = graded positives (score 1–3) plus hard negatives (score 0).

Aspect	corpus	train q	silver q	gold q	train qrels	silver qrels	gold qrels
emotion（情感）	78,758	5,881	1,510	100	142,931	36,783	2,100
intent（意图）	75,631	5,578	1,442	100	130,459	33,752	2,100
theme（主题）	83,072	6,239	1,575	100	157,574	39,736	2,100
thought（思想）	69,039	5,130	1,325	100	126,139	32,500	2,100

Totals: queries ≈ 29,080 · qrels ≈ 708,274

Configs

Each aspect has its own corpus (candidate pools differ by dimension).

Hub Dataset Viewer default: queries_emotion (tabs: train / silver / gold).
Document libraries use split name corpus (not test).

Config pattern	Splits	Description
corpus_emotion … corpus_thought	corpus	Document library for that aspect
queries_emotion … queries_thought	train / silver / gold	Queries
qrels_emotion … qrels_thought	train / silver / gold	Labels (score 0–3; 0=hard negative)

All configs

Aspect	Configs
emotion	corpus_emotion, queries_emotion, qrels_emotion
intent	corpus_intent, queries_intent, qrels_intent
theme	corpus_theme, queries_theme, qrels_theme
thought	corpus_thought, queries_thought, qrels_thought

Relevance scale

Score	Meaning
3	Strong match (incl. source-poem self pair)
2	Moderate match
1	Weak match
0	Hard negative (pooled distractor)

How to load

from datasets import load_dataset

corpus = load_dataset("PoetryMTEB/ClassicalPoetryRetrieval", "corpus_emotion", split="corpus")
queries = load_dataset("PoetryMTEB/ClassicalPoetryRetrieval", "queries_emotion", split="gold")
qrels = load_dataset("PoetryMTEB/ClassicalPoetryRetrieval", "qrels_emotion", split="gold")

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

q_silver = load_dataset("PoetryMTEB/ClassicalPoetryRetrieval", "queries_theme", split="silver")
q_train = load_dataset("PoetryMTEB/ClassicalPoetryRetrieval", "queries_intent", split="train")
# DatasetDict for all query splits:
qs = load_dataset("PoetryMTEB/ClassicalPoetryRetrieval", "queries_emotion")  # train / silver / gold

Construction (summary)

1. Per-aspect candidate pooling + two-round labeling (binary then 4-scale).
2. Cluster-stratified ~8:2 train / test; gold ≈ 100 held-out agreement queries.
3. Gold qrels from human RAA (crowdsourcing + adjudication).
4. Unified PoetryMTEB packaging (this repo).

Annotation prompts

Per-dimension Chinese 4-scale (0–3) labeling prompts are under prompts/:

Path	Content
prompts/prompt_specs.json	Structured ZH rubrics for emotion / intent / theme / thought
prompts/<dim>/system_zh.txt	System prompt
prompts/<dim>/user_template_zh.md	User prompt template
prompts/prompts.py	Source used by the labeling pipeline

<dim> ∈ {emotion, intent, theme, thought}.
Pipeline source: poetry_llm_labels_4scale_full_bundle/scripts/verify_pool_llm_agreement.py
(label_pool_llm_4scale.py imports the same prompt builders). Labeling language: zh.

Citation

@misc{classical_poetry_retrieval_poetrymteb,
  title        = {{Classical Poetry Retrieval}: Multi-Aspect Graded
                  Retrieval for Classical Chinese Poetry},
  author       = {{PoetryMTEB Contributors}},
  year         = {2026},
  version      = {1.2.0},
  publisher    = {{Hugging Face}},
  url          = {{https://huggingface.co/datasets/PoetryMTEB/ClassicalPoetryRetrieval}},
  license      = {{CC-BY-NC-SA-4.0}},
  note         = {{PoetryMTEB BEIR-style retrieval; aspects emotion,
                  intent, theme, thought; graded relevance 1--3;
                  splits train / silver / gold}}
}

License

CC BY-NC-SA 4.0.