CCPM — Pair Classification (PoetryMTEB)

Chinese Classical Poetry Matching Dataset (CCPM), reformatted as an MTEB-style Pair Classification task for PoetryMTEB embedding evaluation.

Upstream source: CCPM (Li et al., 2021). Original task: given a modern Chinese description, select the matching classical poetry line from four candidates.

---

Dataset Card

Item	Description
Source	CCPM — Chinese Classical Poetry Matching Dataset (Li et al., 2021)
Languages	Chinese (zh): modern Chinese description ↔ classical poetry line
Size	108,872 labeled pairs (train 87,112; validation 10,880; test 10,880). Derived from 21,778 / 2,720 / 2,720 original MCQ instances (×4 choices)
Label type	Binary pair label: labels=1 (semantic match), labels=0 (hard negative)
Splits	train ← original train; validation / test ← original validation (see note below)
Construction	Expand each 4-way MCQ into 4 pairs (sentence1=translation, sentence2=choice)
License	CC BY-NC 4.0 (PoetryMTEB redistribution; please cite original authors)
Evaluation metrics	MTEB Pair Classification: Average Precision (AP), accuracy / F1 at best similarity threshold (cosine / Euclidean / Manhattan). Original CCPM MCQ metric: Accuracy

Split note: The public test_public.jsonl in the original release has no gold answers. For Pair Classification we use the original validation set as both validation and test (identical labeled pairs) so MTEB-style evaluators that read test can run. Do not treat train→test leakage across these two identical eval splits.

---

Features

Field	Type	Description
id	string	Unique pair id
sentence1	string	Modern Chinese description (translation)
sentence2	string	Classical poetry line candidate (choice)
labels	int64	1 = matching pair; 0 = non-matching (hard negative)

Schema aligns with MTEB Pair Classification (sentence1, sentence2, labels).

---

Construction method

1. Upstream CCPM builds each instance from classical–modern parallel poetry: the gold line is the correct choice; three distractors are retrieved as similar lines from a classical poetry corpus (hard negatives).
2. This release converts each MCQ instance into four binary pairs:

for choice_i in choices:
    sentence1 = translation
    sentence2 = choice_i
    labels    = 1 if i == answer else 0

Class balance per split: 1 positive : 3 negatives.

---

Size by split

Split	#pairs	#positive (labels=1)	#negative (labels=0)	Origin
train	87,112	21,778	65,334	original train (21,778×4)
validation	10,880	2,720	8,160	original valid (2,720×4)
test	10,880	2,720	8,160	same as validation (see note)

---

Evaluation metrics

PoetryMTEB / MTEB Pair Classification

Metric	Role
Average Precision (AP)	Ranking quality of pair similarity scores
Accuracy / F1	Binary decision at best threshold over cosine / distance scores

Original CCPM (4-way matching) used Accuracy over candidate indices.

---

How to load

from datasets import load_dataset

ds = load_dataset("PoetryMTEB/CCPM")
print(ds)
print(ds["test"][0])
# {'id': '...', 'sentence1': '...', 'sentence2': '...', 'labels': 0 or 1}

---

License

Distributed under CC BY-NC 4.0.
Please cite the original CCPM paper when using this data.

---

Citation

@article{li2021CCPM,
  title = {CCPM: A Chinese Classical Poetry Matching Dataset},
  author = {Li, Wenhao and Qi, Fanchao and Sun, Maosong and Yi, Xiaoyuan and Zhang, Jiarui},
  journal = {arXiv preprint arXiv:2106.01979},
  year = {2021},
  url = {https://arxiv.org/abs/2106.01979}
}

This Hub dataset: PoetryMTEB/CCPM (Pair Classification packaging for PoetryMTEB).