romanized_bangla — Dataset Card

Repository / id: sk-community/romanized_bangla Derived from: wikimedia/wikipedia subset 20231101.bn (Bangla Wikipedia dump)

---

1. Short description

A romanized (Latin-script) version of Bangla Wikipedia text derived from the wikimedia/wikipedia dataset (subset 20231101.bn). The original dataset contained whole-article entries; you split article paragraphs into individual rows and transliterated Bangla text into a romanized representation using a Python script. The processed dataset contains 975,215 rows (paragraph-level documents) in romanized Bangla.

This dataset is intended for tasks such as: language modeling, back-transliteration, transliteration normalization, pretraining/finetuning language models that support Romanized Bangla, and dataset augmentation for downstream Bangla NLP tasks.

---

2. Dataset summary

• Original source: wikimedia/wikipedia — subset 20231101.bn (Bangla).
• Original size (articles): ~143,069 rows (article-level).
• Processed size (paragraph-level, romanized): 975,215 rows (as uploaded to sk-community/romanized_bangla).
• Languages: Romanized Bangla (Latin script).
• Modality: Text

---

• Transliteration script: https://github.com/sk-research-community/multilingual-transliterator-llm-training/blob/main/src/rule_based_bengali_transliterate.py

---

3. Recommended usage

• Tasks: language modeling, transliteration/back-transliteration, normalization, data augmentation, pretraining/finetuning transformer models that should accept romanized Bangla.
• Load example:

from datasets import load_dataset

ds = load_dataset("sk-community/romanized_bangla", split="train")
print(next(iter(ds))["text"])  # prints a romanized paragraph

• Example preprocessing for training: tokenization with an appropriate tokenizer (Byte-Pair, SentencePiece, or multilingual tokenizers such as XLM-R based tokenizers) and typical ML preprocessing pipelines.

---

4. Train/validation/test splits

Currently the dataset is provided as a single train split. If you plan to publish splits for downstream tasks, consider adding train/validation/test splits and provide reproducible splitting code (e.g., deterministic by hashing id or using datasets train_test_split).

---

5. Licensing and terms of use

This dataset is distributed under the MIT License. Please also check the licenses of the individual source datasets before use.

6. Citation

@article{gharami2025indotranslit,
  title={Modeling Romanized Hindi and Bengali: Dataset Creation and Multilingual LLM Integration},
  author={Kanchon Gharami and Quazi Sarwar Muhtaseem and Deepti Gupta and Lavanya Elluri and Shafika Showkat Moni},
  year={2025}
}

---

7. Maintainers / Contact

• Maintainer: sk-community

---