Oni279/BanglaMedQA

---
license: cc-by-sa-4.0
---

# Dataset Card for BanglaMedQA and BanglaMMedBench

This dataset introduces **BanglaMedQA** and **BanglaMMedBench**, the first large-scale Bangla biomedical Multiple Choice Question (MCQ) datasets designed to evaluate reasoning and retrieval-based methods such as Retrieval-Augmented Generation (RAG) for Bangla Question Answering.

## Dataset Details

### Dataset Description

**BanglaMedQA** consists of 1,000 MCQs collected from authentic Bangladeshi medical admission exams (MBBS, BDS, AFMC) between 1990 and 2024, each with reasoning for the correct answer.
**BanglaMMedBench** contains 1,000 scenario-based medical questions translated from the English MMedBench dataset using Gemini-1.5-Flash, reviewed by medical experts for domain and linguistic accuracy.

- **Curated by:** Department of Computer Science and Engineering, Islamic University of Technology (IUT), Bangladesh
- **Funded by [optional]:** IUT Research Fund
- **Shared by [optional]:** Sadia Sultana, Saiyma Sittul Muna, Mosammat Zannatul Samarukh, Ajwad Abrar, Tareque Mohmud Chowdhury
- **Language(s) (NLP):** Bangla
- **License:** Not given

### Dataset Sources [optional]

- **Repository:** [https://huggingface.co/datasets/ajwad-abrar/BanglaMedQA](https://huggingface.co/datasets/ajwad-abrar/BanglaMedQA)
- **Paper [optional]:** [BanglaMedQA and BanglaMMedBench: Evaluating Retrieval-Augmented Generation Strategies for Bangla Biomedical Question Answering (arXiv:2511.04560v1)](https://arxiv.org/abs/2511.04560v1)
- **Demo [optional]:** Not given

## Uses

### Direct Use

- Benchmarking Bangla biomedical QA systems
- Evaluating Retrieval-Augmented Generation (RAG) and reasoning strategies
- Studying translation and reasoning robustness across English–Bangla in medical domains

### Out-of-Scope Use

- Not suitable for clinical or diagnostic use
- Not to be used for real patient advice or decision-making
- Should not replace professional medical education or assessment materials

## Dataset Structure

Each dataset contains 1,000 entries with the following fields:

- `question`: Medical question in Bangla
- `options`: Four options labeled A, B, C, D
- `answer`: Correct option label
- `rationale`: Explanation for the correct answer

No predefined train/test/validation split is provided; users may define splits as needed.

## Dataset Creation

### Curation Rationale

The datasets were created to fill the gap in Bangla biomedical QA resources and to evaluate retrieval-augmented reasoning performance in low-resource languages for educational and AI research purposes.

### Source Data

#### Data Collection and Processing

- **BanglaMedQA:**
- Sourced from 34 years of official Bangladeshi medical admission exams.
- Removed ambiguous or duplicate questions.
- Standardized multiple-choice formats and corrected inconsistencies.
- **BanglaMMedBench:**
- Translated from English MMedBench using Gemini-1.5-Flash.
- Manually verified for medical and linguistic accuracy by domain experts.
- Formatting and alignment corrected post-translation.

#### Who are the source data producers?

- Medical admission exam boards of Bangladesh (MBBS, BDS, AFMC)
- Original MMedBench dataset creators

### Annotations [optional]

#### Annotation process

Each question includes a verified rationale explaining the correct answer.
Quality control was performed by IUT researchers and reviewed by medical experts for correctness and clarity.

#### Who are the annotators?

Undergraduate and graduate students at IUT under faculty supervision, with medical domain experts validating the rationales.

#### Personal and Sensitive Information

No personal or sensitive data are included.
The datasets contain only educational biomedical content.

## Bias, Risks, and Limitations

- Translation from English may cause minor semantic drift.
- OCR-based textbook corpus may introduce small transcription errors.
- Dataset size (2,000 questions total) may limit generalization to broader medical subfields.
- Rationale style and vocabulary may reflect academic exam phrasing.

### Recommendations

Use these datasets for benchmarking and research only.
Cross-verify terminology when building multilingual or fine-tuned models.
Do not apply outputs in real-world medical or diagnostic settings.

## Citation [optional]

**BibTeX:**
```bibtex
@article{Sultana2025BanglaMedQA,
title={BanglaMedQA and BanglaMMedBench: Evaluating Retrieval-Augmented Generation Strategies for Bangla Biomedical Question Answering},
author={Sultana, Sadia and Muna, Saiyma Sittul and Samarukh, Mosammat Zannatul and Abrar, Ajwad and Chowdhury, Tareque Mohmud},
year={2025},
journal={arXiv preprint arXiv:2511.04560},
url={https://arxiv.org/abs/2511.04560v1}
}