README.md

metadata

task_categories:
  - text2text-generation
language:
  - bn
tags:
  - bangla
  - punctuation-restoration
  - adversarial-examples
  - seq2seq
  - nlp
pretty_name: Bangla Punctuation Restoration (Adversarial Attack)
size_categories:
  - 10K<n<100K

ha-pr-bn-aminul-attack

Dataset Name: ha-pr-bn-aminul-attack
Language: Bengali (bn)
Task: Punctuation Restoration (Text-to-Text Generation)
Size: 11,113 conversations
Format: Human-assistant conversations with adversarially perturbed inputs

Dataset Description

This dataset extends ha-pr-bn-aminul-generated with adversarial attacks on the unpunctuated inputs. It simulates noisy or corrupted Bangla text, often produced by OCR, ASR, or user typos. The assistant still returns the original, correct punctuated version.

Dataset Summary

• Language: Bengali (bn)
• Task: Punctuation Restoration (Text-to-Text Generation)
• Size: 11,113 adversarial examples
• License: CC BY 4.0
• Purpose: Test the robustness of punctuation models under noisy input conditions

Dataset Structure

Data Instances

{
  "input": "শুধু বর্ষা নয় শীত বা গ্রীষ্মেও বেড়ে যাচ্ছে তাপ",
  "target": "শুধু বর্ষা নয়, শীত বা গ্রীষ্মেও বেড়ে যাচ্ছে তাপ।"
}

Usage

from datasets import load_dataset

dataset = load_dataset("aminul/ha-pr-bn-aminul-attack")
print(dataset["train"][0])

Dataset Creation

Attack Methods

Generated using TextAttack techniques:

• Word-level swaps
• Misspellings
• OCR-style distortions
• Transliteration and ASR-style noise

Processing

1. Perturb the input sentence
2. Preserve original target (punctuated) text