---
language:
  - bn
pipeline_tag: token-classification
library_name: transformers
base_model: csebuetnlp/banglabert
base_model_relation: finetune
license: other
datasets:
  - abdullahalmunem/ha-pr-bn-munem-generated
tags:
  - bengali
  - bangla
  - punctuation-restoration
  - asr-post-processing
  - banglaprcorpus
---

# Bangla Punctuation Restoration Extended v3

An experimental ten-class Bengali punctuation-restoration model fine-tuned from
the local punctuation-v2 checkpoint, whose encoder originated from
[`csebuetnlp/banglabert`](https://huggingface.co/csebuetnlp/banglabert).

Compared with v2, this checkpoint adds hyphen (`-`) and Bengali visarga (`ঃ`).
It was trained with data from Munem's generated punctuation corpus,
BanglaPRCorpus, and replay data used to reduce forgetting of the original task.

## Labels

| ID | Label |
|---:|:---|
| 0 | `O` (no punctuation) |
| 1 | `,` |
| 2 | `।` |
| 3 | `?` |
| 4 | `!` |
| 5 | `;` |
| 6 | `:` |
| 7 | `…` |
| 8 | `-` |
| 9 | `ঃ` |

## Evaluation summary

The final independent evaluations showed that performance remains uneven:

| Test set | Overall macro F1 | Punctuation macro F1 |
|:---|---:|---:|
| Munem generated | 0.447 | 0.389 |
| BanglaPRCorpus | 0.396 | 0.331 |
| Original test | 0.473 | 0.416 |

The new-class F1 scores were approximately `0.385`/`0.500` on the Munem test
for hyphen/visarga and `0.445`/`0.136` on BanglaPRCorpus. This model should be
considered experimental, not a universal improvement over v2.

## Usage

### Recommended: Python package

```bash
pip install bangla-punctuation
```

```python
from bangla_punctuation import BanglaPunctuator

punctuator = BanglaPunctuator(model="extended", 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](https://pypi.org/project/bangla-punctuation/)
for the complete Python and ASR integration documentation.

### Direct Transformers loading

```python
from transformers import AutoModelForTokenClassification, AutoTokenizer

model_id = "HasinManjare/bangla-punctuation-extended-v3"
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](https://huggingface.co/HasinManjare)).

- Stable model: [`bangla-punctuation-v2`](https://huggingface.co/HasinManjare/bangla-punctuation-v2)
- Extended model: [`bangla-punctuation-extended-v3`](https://huggingface.co/HasinManjare/bangla-punctuation-extended-v3)
- Python library: [`bangla-punctuation`](https://pypi.org/project/bangla-punctuation/)

## Data and attribution

- Encoder: [`csebuetnlp/banglabert`](https://huggingface.co/csebuetnlp/banglabert),
  Bhattacharjee et al., NAACL 2022.
- Generated corpus:
  [`abdullahalmunem/ha-pr-bn-munem-generated`](https://huggingface.co/datasets/abdullahalmunem/ha-pr-bn-munem-generated),
  published under Apache-2.0.
- BanglaPRCorpus: Mehedi Hasan Bijoy et al., "Advancing Bangla Punctuation
  Restoration by a Monolingual Transformer-Based Method and a Large-Scale
  Corpus," BLP/EMNLP 2023. Its GitHub repository displays an MIT license.

## Citing this model

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

```bibtex
@misc{pitom2026banglapunctuationextended,
  author = {Manjara Hasin Al Pitom},
  title = {Bangla Punctuation Restoration Extended v3},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/HasinManjare/bangla-punctuation-extended-v3}}
}
```

## 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. Weak and
new punctuation classes have limited accuracy, and performance may degrade on
noisy ASR output, dialects, code-switching, or unfamiliar domains.