README.md

---
license: apache-2.0
task_categories:
- token-classification
- text-classification
tags:
- code
---

# Bangla Punctuation Restoration Dataset

A merged, high-quality Bangla dataset for **punctuation restoration**, formatted as **instruction-tuning conversation pairs**.  
The dataset is suitable for fine-tuning Large Language Models (LLMs) and sequence models to restore punctuation in Bangla text.

---

## Dataset Summary

- **Language**: Bengali (Bangla)
- **Task**: Punctuation Restoration
- **Format**: JSONL (instruction-style conversations)
- **Max chunk length**: ~256 characters
- **Punctuation covered**:  
  `। ! ? , ; : -`

Each example contains:
- An **unpunctuated Bangla text** as input
- A **punctuated Bangla text** as output

---

## Data Format

Each dataset entry follows the structure below:

```json
{
  "conversations": [
    {
      "from": "human",
      "value": "এটি একটি উদাহরণ বাক্য যেখানে কোন বিরামচিহ্ন নেই"
    },
    {
      "from": "gpt",
      "value": "এটি একটি উদাহরণ বাক্য, যেখানে কোনো বিরামচিহ্ন নেই।"
    }
  ],
  "source": "source_name",
  "score": 7.83
}
````

### Field Description

| Field           | Description                                        |
| --------------- | -------------------------------------------------- |
| `conversations` | Instruction-style input–output pair                |
| `source`        | Origin of the sample                               |
| `score`         | Random float (4.0–10.0), placeholder quality score |

---

## Data Sources

This dataset is created by merging and processing three main sources:

### 1. Badhon/BanglaQuranPunctuationDataset (Hugging Face)

* Bangla Quran text with accurate punctuation
* High grammatical and punctuation consistency

### 2. hishab/hishab-pr-bn-v1 (Hugging Face)

* Bangla punctuation restoration dataset
* Diverse sentence structures and punctuation styles

### 3. Web-Collected Bangla Text (Custom Scraper)

High-quality Bangla paragraphs collected from:

* Bengali Wikipedia
* Prothom Alo (`prothomalo.com`)
* BBC Bangla (`bbc.com/bengali`)
* Bangla blogs:

  * somewhereinblog.net
  * sachalayatan.com
  * amarblog.com

---

## Data Processing

The following preprocessing steps were applied:

### Cleaning

* Preserved Bangla punctuation (`। ! ? , ; : -`)
* Removed encoding artifacts and noisy symbols
* Ensured Bangla language dominance

### Chunking

* Text split into chunks of ≤256 characters
* Sentence-boundary–aware splitting
* Avoided mid-sentence truncation

### Quality Filtering

* Minimum and maximum length thresholds
* Required presence of natural punctuation
* Removed malformed or low-quality samples

### Deduplication

* Removed duplicates using punctuated-text matching
* Improved diversity and reduced redundancy

---

## Intended Uses

* Fine-tuning LLMs for Bangla punctuation restoration
* Training instruction-following Bangla NLP models
* Sequence-to-sequence or token-classification approaches
* ASR post-processing pipelines (speech → text → punctuation)