SahittoCategoryClassification

Single-label Bangla poem category / genre classification for PoetryMTEB embedding evaluation, derived from the Bangla Poem Dataset (SAHITTO).

Dataset Card

Item	Description
Source	Bangla Poem Dataset (Mendeley), DOI 10.17632/zgmrk5m566.2; poems collected from banglarkobita.com
Languages	Bangla / Bengali (bn)
Size	2207 poems total; train=1763 (79.9%); validation=222 (10.1%); test=222 (10.1%)
Classes	11 poem categories (题材/主题)
Splits	Stratified by category ≈ 80% / 10% / 10% (train / validation / test), seed=42
License	CC BY 4.0 (same as upstream)
Evaluation metrics	Classification on embeddings: accuracy, macro/weighted F1

---

Split statistics

Split	#poems	ratio
train	1763	79.88%
validation	222	10.06%
test	222	10.06%
total	2207	100%

Per-category counts by split

id	label_name	Bangla	train	validation	test	total
0	Miscellaneous	বিবিধ কবিতা	649	81	81	811
1	Love	প্রেমের কবিতা	400	50	50	500
2	Metaphor	রূপক কবিতা	252	32	32	316
3	Separation	বিরহের কবিতা	105	13	13	131
4	Humanity	মানবতাবাদী কবিতা	63	8	8	79
5	Children	ছোটদের ছড়া-কবিতা	62	8	8	78
6	Patriotic	দেশাত্মবোধক কবিতা	60	8	8	76
7	Policy	নীতি কবিতা	53	7	7	67
8	Religious	ধর্মীয় কবিতা	50	6	6	62
9	Nature	প্রকৃতির কবিতা	43	6	6	55
10	War	যুদ্ধের কবিতা	26	3	3	32
	Sum		1763	222	222	2207

Codebook: label_taxonomy.json.

---

Features

Field	Type	Description
id	string	Example id
title	string	Poem title
author	string	Writer name (Bangla)
poem	string	Poem body
label	int64	Category index (0 … C−1)
label_name	string	Category name in English (upstream label)

---

Construction method

1. Load SAHITTO.ods (sheet Poems: title, poem, writer, label).
2. Keep all 11 category labels; map to contiguous label ids (frequency-descending).
3. Stratified train / validation / test split by category (each class appears in all three splits).

---

How to load

from datasets import load_dataset

ds = load_dataset("PoetryMTEB/SahittoCategoryClassification")
print(ds["test"][0]["label_name"], ds["test"][0]["title"])

For embedding evaluation, encode poem (optionally prepend title), fit a classifier on train, optionally tune on validation, score on test.

---

License

Distributed under Creative Commons Attribution 4.0 International (CC BY 4.0), consistent with the upstream Mendeley release.

---

Citation / provenance

Please cite the Bangla poem genre classification paper and the Mendeley dataset:

@INPROCEEDINGS{10174592,
  author={Pasha, Syed Tangim and Islam, Ashraful and Rahman, Mohammed Masudur and Ahmed, Eshtiak and Ahmed Foysal, Md. Ferdouse and Zahangir Alam, Md},
  booktitle={2023 IEEE World AI IoT Congress (AIIoT)}, 
  title={Genre Classification of Bangla Poem Using Machine Learning and Deep Learning Techniques}, 
  year={2023},
  pages={0236-0242},
  doi={10.1109/AIIoT58121.2023.10174592}
}

Dataset citation:

@misc{pasha2022banglapoem,
  author = {Pasha, Syed Tangim},
  title = {Bangla Poem Dataset},
  year = {2022},
  publisher = {Mendeley Data},
  version = {V2},
  doi = {10.17632/zgmrk5m566.2},
  url = {https://data.mendeley.com/datasets/zgmrk5m566/2}
}

• Mendeley: https://data.mendeley.com/datasets/zgmrk5m566/2
• Paper: https://doi.org/10.1109/AIIoT58121.2023.10174592
• This Hub packaging: PoetryMTEB/SahittoCategoryClassification