BangladeshiVQA

A culturally grounded Bangla Visual Question Answering benchmark.

BangladeshiVQA is a native Bangla VQA benchmark of 2,068 Bangladeshi images and 7,038 open-ended question–answer pairs, organized into three cognitive levels and seven image categories. To our knowledge it is the first Bangla VQA dataset with dedicated in-image Bangla scene-text (OCR) questions, and the first to split strictly by image ID to prevent train/test leakage.

This Hugging Face repository hosts the full dataset (all images + all annotations + the splits). Code, the SmolVLM2 fine-tuning configuration, and a small preview sample live in the accompanying GitHub repository — see the paper for the link.

• 📄 Paper: forthcoming — link will be added upon publication
• 💻 Code (GitHub): https://github.com/tanim494/BangladeshiVQA
• ⚖️ License: Annotations under CC BY-NC 4.0; images remain under their original ownership and are provided for non-commercial academic research only.

Note on naming. During development this dataset was called ChitroSamagra. It has been renamed BangladeshiVQA; older drafts or artifacts may still use the former name.

Files

images/                       # 2,068 full-resolution JPEGs, named CS_0001.jpg … 
annotations/
├── qa_pairs.json             # all 7,038 QA pairs (flat records)
└── splits/
    ├── train.json            # 5,628 pairs  (1,653 images)
    ├── val.json              #   704 pairs  (  211 images)
    └── test.json             #   706 pairs  (  204 images)

Splits are disjoint by image ID — no image appears in more than one split.

Record format

Each entry in qa_pairs.json and the split files is a flat record:

{
  "image_id": "CS_0001",
  "q_type": "Counting",
  "question": "কয়টি সিএনজি অটোরিকশা দেখা যাচ্ছে?",
  "answer": "২ টি",
  "level": "Level 1"
}

image_id matches the image filename (CS_0001 → images/CS_0001.jpg). q_type ∈ {Existence, Counting, Attribute, Reasoning, OCR}; level ∈ {Level 1, Level 2, Level 3}, derived from the question type.

Dataset at a glance

Images	2,068
QA pairs	7,038
Splits (by image ID)	train 5,628 · val 704 · test 706
Cognitive levels	3 (Perception, Representation, Cognition)
Question types	Existence, Counting, Attribute, Reasoning, OCR
Image categories	7
Language	Bangla (Bengali)
Unique answers	1,524

Question taxonomy

Level	Question type	Count
Level 1 — Perception		3,245
	Existence	2,069
	Counting	1,176
Level 2 — Representation		3,280
	Attribute	3,280
Level 3 — Cognition		513
	Reasoning	375
	OCR	138

Loading

from huggingface_hub import hf_hub_download, snapshot_download
import json

# Annotations
path = hf_hub_download(repo_id="tanim494/BangladeshiVQA", filename="annotations/qa_pairs.json",
                       repo_type="dataset")
qa = json.load(open(path, encoding="utf-8"))
print(len(qa), "QA pairs")   # 7038

# Full dataset (images + annotations)
local = snapshot_download(repo_id="tanim494/BangladeshiVQA", repo_type="dataset",
                          local_dir="./BangladeshiVQA")

Intended use & limitations

Non-commercial academic research — evaluating and fine-tuning vision-language models on Bangladeshi geo-cultural and tiered multimodal reasoning. Out of scope: commercial use; any person-identification, biometric, or surveillance application; treating the dataset as representative of all of Bangladesh (it is a curated cultural sample).

Privacy: released images were verified with ExifTool to contain no EXIF/GPS metadata and were renamed to prevent recovery of source filenames. See the GitHub DATASHEET.md and NOTICE.md for the full ethics, consent, and takedown details.

Citation

The accompanying paper is forthcoming; its citation will be added here on publication. In the meantime, please cite the dataset:

@misc{bangladeshivqa_dataset,
  title     = {BangladeshiVQA},
  author    = {Ahsanul Karim Tanim and Durjoy Saha},
  year      = {2026},
  publisher = {Hugging Face},
  doi       = {10.57967/hf/10089},
  url       = {https://huggingface.co/datasets/tanim494/BangladeshiVQA}
}