ChitraMiti-12.8k is currently under review for MathNLP 2026. To access the benchmark splits, questions, and diagram assets, please provide your research affiliation and details below. Access is granted for non-commercial academic research.

ChitraMiti: Benchmarking Visual Grounding and Modality Reliance in Bengali Geometric Reasoning 📐

Author & Maintainer

Khan Raiyan Ibne Reza
Department of Electrical & Computer Engineering, North South University
📧 Email: raiyan.reza@northsouth.edu
🔗 GitHub: ChitraMiti Repository

📌 Overview

ChitraMiti is the first comprehensive, expert-verified multimodal geometric reasoning benchmark designed specifically for the low-resource Bengali language.

Aligned with the Bangladesh NCTB (National Curriculum and Textbook Board) secondary mathematics curriculum (Classes 6–10), ChitraMiti evaluates Vision-Language Models (VLMs) on their visual grounding capabilities, spatial deduction, and reliance across modalities.

🌟 Key Highlights

Largest Bengali Multimodal Geometry Corpus:
- ChitraMiti-12.8k (Synthetic): 12,874 diagram-question pairs partitioned into disjoint training (11,231), validation (643), and test (1,000) splits.
- NCTB-500 (Real-World Textbook Benchmark): 500 diagram problems extracted directly from official NCTB textbook pages scanned at 300 DPI, completely uncorrupted by automated diagram OCR.
15-Attribute Structural Descriptions:
- Every diagram is grounded by a rigorous 15-attribute topological description generated and refined to capture all spatial, numerical, and symbolic attributes.
Three-Phase Evaluation Protocol:
- Phase A (Diagram Only): Diagram + Question.
- Phase B (Diagram + Description): Diagram + Description + Question.
- Phase C (Description Only): Description + Question.
- Tests whether VLMs genuinely reason from diagram images or bypass the visual modality when text descriptions are supplied (evaluated with paired Two One-Sided Tests — TOST).

📊 Dataset Structure & Statistics

1. Corpus Summary

Configuration	Split	Records	Modality	Primary Fields
`default` (Fine-tuning)	`train`	11,231	Image + Text	`id`, `image`, `bangla_question`, `answer`, `visual_description`, `messages`
	`validation`	643	Image + Text	`id`, `image`, `bangla_question`, `answer`, `visual_description`, `messages`
	`test`	1,000	Image + Text	`id`, `image`, `bangla_question`, `answer`, `visual_description`, `messages`
`evaluation_1k`	`test`	1,000	Image + Text	`node_id`, `image`, `prompt`, `answer`, `visual_description`, `extraction_method`
`nctb_500`	`test`	500	Real Image + Text	`node_id`, `image`, `prompt`, `answer`, `visual_description`, `chapter`, `difficulty`, `type`
`corpus_index`	`train`	12,874	Image + Text	`node_id`, `image`, `visual_description`, `geometric_type`, `concepts`

2. NCTB-500 Curriculum Breakdown

Chapter / Topic	Bengali	Count	% of Benchmark
Trigonometry	ত্রিকোণমিতি	193	38.6%
General Geometry	সাধারণ জ্যামিতি	116	23.2%
Shapes	জ্যামিতিক আকৃতি	50	10.0%
Triangles	ত্রিভুজ	40	8.0%
Circles	বৃত্ত	40	8.0%
Similarity	সদৃশতা	31	6.2%
Mensuration	ক্ষেত্রমিতি	30	6.0%
Total		500	100.0%

Difficulty Breakdown: Hard (78.6%), Medium (19.0%), Easy (2.4%).

🧩 Instance Example

{
  "id": "CM_train_0001",
  "image": "<PIL.Image.Image image mode=RGB size=512x512>",
  "bangla_question": "চিত্রে প্রদর্শিত সমকোণী ত্রিভুজ ABC-এর অতিভুজ AC-এর দৈর্ঘ্য কত?",
  "answer": "৫",
  "visual_description": "চিত্রটিতে একটি সমকোণী ত্রিভুজ ABC রয়েছে যার কোণ ∠B = ৯০°। উল্লম্ব বাহু AB = ৩ একক এবং অনুভূমিক বাহু BC = ৪ একক...",
  "messages": [
    {"role": "system", "content": "তুমি একজন গণিত বিশেষজ্ঞ..."},
    {"role": "user", "content": [{"type": "image"}, {"type": "text", "text": "..."}]},
    {"role": "assistant", "content": [{"type": "text", "text": "..."}]}
  ]
}

🚀 Quick Start & Usage

Because this dataset contains embedded images, you can load and inspect it directly with the Hugging Face datasets library once access is approved:

from datasets import load_dataset

# 1. Load the primary multimodal fine-tuning splits
dataset = load_dataset("RaiyanKhaan/ChitraMiti-12.8k")
print(dataset)
# Output:
# DatasetDict({
#     train: Dataset({features: ['id', 'image', 'bangla_question', 'answer', 'visual_description', 'messages'], num_rows: 11231}),
#     validation: Dataset({features: [...], num_rows: 643}),
#     test: Dataset({features: [...], num_rows: 1000})
# })

# Access the first training instance
sample = dataset["train"][0]
image = sample["image"]  # Returns a PIL Image directly
print("Question:", sample["bangla_question"])
print("Answer:", sample["answer"])

# 2. Load the synthetic 1k Evaluation Benchmark
eval_benchmark = load_dataset("RaiyanKhaan/ChitraMiti-12.8k", "evaluation_1k", split="test")

# 3. Load the Real-World NCTB-500 Textbook Benchmark
nctb_benchmark = load_dataset("RaiyanKhaan/ChitraMiti-12.8k", "nctb_500", split="test")
print(f"Loaded {len(nctb_benchmark)} real textbook geometry items.")

🔒 Access and Download Request

This repository is Gated to maintain an academic record of downloads while the paper is under review:

To request access, please click "Request Access" above and provide your name, university affiliation, and research purpose.
Requests for academic, educational, and non-commercial research use are reviewed and approved regularly.

⚖️ Ethics, Licensing, and Usage Terms

ChitraMiti-12.8k (Synthetic Benchmark):
- Annotations, questions, answers, and 15-attribute structured visual descriptions: Licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).
- Synthetic diagrams: Inherits the DeepVision-103k planar subset licensing.
NCTB-500 (Real-World Benchmark):
- Derived annotations, questions, and structured metadata: Licensed under CC BY 4.0.
- Textbook diagram crops: The underlying educational assets belong to the National Curriculum and Textbook Board (NCTB), Bangladesh. High-resolution crops are made available strictly for non-commercial academic research under Fair Use for reproducible model evaluation. Full textbook pages are excluded.

📖 Citation

If you find ChitraMiti useful for your research, please cite:

@article{reza2026chitramiti,
  title={ChitraMiti: Benchmarking Visual Grounding and Modality Reliance in Bengali Geometric Reasoning},
  author={Reza, Khan Raiyan Ibne, Sanjana Aktar Maria, Dr. Sumaiya Tabassum Nimi, and Dr. Md Adnan Arefeen},
  journal={N/A},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/RaiyanKhaan/ChitraMiti-12.8k}
}

Downloads last month: 16

Total file size:

509 MB