Badhon/ha-pr-bn-Badhon-generated

Badhon PR Bengali Generated Dataset

Dataset Description

This dataset contains Bengali text punctuation restoration data in a conversational format, generated using GPT on Wikipedia content. It is designed for training and evaluating models to restore punctuation in Bengali text. Each conversation consists of a human providing unpunctuated Bengali text and a GPT assistant providing the same text with proper punctuation restored.

Dataset Summary

• Language: Bengali (bn)
• Task: Punctuation Restoration (Text-to-Text Generation)
• Size: 3,827 conversations
• Format: Human-assistant conversations for punctuation restoration
• License: Apache 2.0

Dataset Structure

Data Instances

Each instance contains a punctuation restoration conversation with metadata:

{
  "conversations": [
    {
      "from": "human",
      "value": "বাংলাদেশের জাতীয় পতাকা সবুজ মাঝে লাল বৃত্ত"
    },
    {
      "from": "gpt",
      "value": "বাংলাদেশের জাতীয় পতাকা সবুজ, মাঝে লাল বৃত্ত।"
    }
  ],
  "source": "wikipedia",
  "score": 8.5
}

Data Fields

• conversations: List of exactly 2 messages:
  • First: from "human" with unpunctuated Bengali text
  • Second: from "gpt" with punctuated version
• source: Origin of the text (e.g., "wikipedia")
• score: Quality score (4.0-10.0 range)

Usage

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("Badhon/ha-pr-bn-Badhon-generated")
train_data = dataset["train"]
print(train_data[0])

Dataset Creation

Source Data

Generated from Bengali Wikipedia articles, processed into punctuation restoration pairs.

Data Processing

• Collected punctuated texts
• Removed punctuation for inputs
• Used GPT to generate/restored punctuated targets
• Assigned quality scores

Considerations for Use

Improves Bengali text processing, readability, and NLP tools.

Additional Information

Licensing

Apache 2.0 – allows research and commercial use with attribution.

Citation

@dataset{ha_pr_bn_badhon_generated,
  title={Badhon PR Bengali Generated Punctuation Restoration Dataset},
  author={Badhon},
  year={2025},
  url={https://huggingface.co/datasets/Badhon/ha-pr-bn-Badhon-generated},
  note={Generated Bengali punctuation restoration dataset with 3,827 examples from Wikipedia},
  license={Apache-2.0}
}