Dataset Card for SQuAD-BN-Passage-to-Question

Dataset Summary

SQuAD-BN-Passage-to-Question is a reformatted and filtered version of the Bangla Question Answering dataset derived from csebuetnlp/squad_bn. The dataset is designed for Bangla text generation and question generation research tasks.

In the original dataset, each context passage is associated with multiple question-answer pairs stored as separate entries. In this modified version:

• All questions associated with the same context passage are grouped together.
• Questions that are irrelevant or weakly related to their corresponding context passages have been removed.

This restructuring improves dataset quality and supports research in:

• Bangla multi-question generation
• Context-aware question generation
• Passage-level instruction tuning
• Bangla NLP low-resource research

---

Dataset Structure

Each dataset entry contains:

• task_id: Unique identifier for each context entry
• context: A Bangla passage of text
• question: A list of Bangla questions associated with the context

Example

{
  "task_id": 1,
  "context": "বাংলা ভাষা দক্ষিণ এশিয়ার একটি প্রধান ভাষা...",
  "question": [
    "বাংলা ভাষা প্রধানত কোথায় ব্যবহৃত হয়?",
    "বাংলা ভাষার উৎপত্তি কোথায়?"
  ]
}

---

Data Splits

The dataset splits directly follow the original csebuetnlp/squad_bn dataset splits:

Split	Source
Train	squad_bn Train Set
Validation	squad_bn Validation Set
Test	squad_bn Test Set

Dataset Statistics

• Language: Bangla (Bengali)
• Source dataset: squad_bn
• Task type: Question Generation / Text Generation
• Data format: JSONL
• Questions per context: Variable (multiple questions grouped together)
• Additional filtering: Irrelevant questions removed

---

Source Dataset

This dataset is derived from:

Bangla SQuAD Dataset (squad_bn)

• Organization: CSE BUET NLP
• Hugging Face Dataset: https://huggingface.co/datasets/csebuetnlp/squad_bn

Modifications from Original Dataset

The following modifications were applied:

1. Grouped multiple questions under a single shared context passage.
2. Removed answer annotations from the original dataset.
3. Removed questions that are irrelevant or weakly related to their corresponding context.
4. Reorganized dataset structure into JSONL format.
5. Added unique task_id identifiers for each context entry.
6. Preserved original dataset split structure.

---

Intended Uses

This dataset is intended for:

• Bangla question generation research
• Instruction tuning for Bangla LLMs
• Text-to-text generation tasks
• Context-based multi-output generation
• Low-resource language NLP research

Recommended Usage

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("Siam0703/SQuAD-BN-Passage-to-Question")

Accessing Data

dataset["train"][0]

Potential Applications

• Bangla educational AI systems
• Context-driven Bangla question synthesis
• Fine-tuning Bangla language models
• Multilingual model benchmarking

---

Limitations

• Answer annotations from the original dataset are not included.
• The dataset inherits biases and limitations from the original squad_bn dataset.
• Manual or rule-based filtering of irrelevant questions may introduce subjective bias.
• Grouped question format may require preprocessing for extractive QA tasks.

---

Ethical Considerations

• The dataset contains human-authored Bangla content collected for educational and research purposes.
• Filtering decisions may introduce unintended bias.
• Models trained on this dataset should be evaluated carefully for fairness and accuracy in Bangla language contexts.

---

Citation

If you use this dataset, please cite the following work:

@misc{bhattacharjee2021banglabert,
      title={BanglaBERT: Combating Embedding Barrier in Multilingual Models for Low-Resource Language Understanding},
      author={Abhik Bhattacharjee and Tahmid Hasan and Kazi Samin and Md Saiful Islam and M. Sohel Rahman and Anindya Iqbal and Rifat Shahriyar},
      year={2021},
      eprint={2101.00204},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

---

Reproducibility Notes

The dataset was generated by regrouping squad_bn entries based on shared context passages and applying filtering to remove irrelevant questions. The original squad_bn dataset is publicly available for reconstruction and comparison.