HasinManjare/bangla-punctuation-v2

Bangla Punctuation Restoration v2

An eight-class Bengali punctuation-restoration model fine-tuned from csebuetnlp/banglabert. It is intended for restoring punctuation in unpunctuated Bengali text, including ASR transcripts.

Labels

ID	Label
0	O (no punctuation)
1	,
2	।
3	?
4	!
5	;
6	:
7	…

Evaluation

On the held-out original test split, punctuation macro F1 was approximately 0.570 and overall macro F1 was approximately 0.623. Performance varies substantially by punctuation class; evaluate the model on data representative of the intended ASR system before production use.

Usage

Recommended: Python package

pip install bangla-punctuation

from bangla_punctuation import BanglaPunctuator

punctuator = BanglaPunctuator(model="base", device="auto")
text = punctuator.restore("আজ আকাশ খুব সুন্দর তুমি কি বাইরে যাবে")
print(text)

The package also supports batch processing with restore_batch(). See bangla-punctuation on PyPI for the complete Python and ASR integration documentation.

Direct Transformers loading

from transformers import AutoModelForTokenClassification, AutoTokenizer

model_id = "HasinManjare/bangla-punctuation-v2"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForTokenClassification.from_pretrained(model_id)

Token-level predictions must be converted back into text using word/subword alignment and the label mapping stored in config.json.

Model developer

Developed and fine-tuned by Manjara Hasin Al Pitom (HasinManjare).

• Stable model: bangla-punctuation-v2
• Extended model: bangla-punctuation-extended-v3
• Python library: bangla-punctuation

Attribution

The encoder is BanglaBERT, introduced by Bhattacharjee et al. (NAACL 2022):

@inproceedings{bhattacharjee-etal-2022-banglabert,
  title = {BanglaBERT: Language Model Pretraining and Benchmarks for Low-Resource Language Understanding Evaluation in Bangla},
  author = {Bhattacharjee, Abhik and Hasan, Tahmid and Ahmad, Wasi and Mubasshir, Kazi Samin and Islam, Md Saiful and Iqbal, Anindya and Rahman, M. Sohel and Shahriyar, Rifat},
  booktitle = {Findings of NAACL},
  year = {2022},
  url = {https://aclanthology.org/2022.findings-naacl.98/}
}

Citing this model

If you use this fine-tuned punctuation-restoration model, please cite both this model and the upstream BanglaBERT work:

@misc{pitom2026banglapunctuation,
  author = {Manjara Hasin Al Pitom},
  title = {Bangla Punctuation Restoration v2},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/HasinManjare/bangla-punctuation-v2}}
}

License and limitations

The upstream csebuetnlp/banglabert release does not currently declare an explicit license in its Hugging Face model card or GitHub repository. Therefore this derived checkpoint is marked license: other; no broader rights to the upstream weights or training data are asserted here. Confirm the applicable permissions with the upstream authors before commercial redistribution or use.

This model may produce incorrect, excessive, or missing punctuation. It should not be treated as a grammatical authority, and performance may degrade on noisy ASR output, dialects, code-switching, or domains unlike its training data.