license: apache-2.0

DiagramQG: A Dataset for Diagram-Driven Course Questions Generation

DiagramQG Dataset

Figure 1: Four different examples of different subjects in DiagramQG dataset.

Figure 2: Domain diversity in DiagramQG where each color corresponds to one discipline.

Overview

DiagramQG is a comprehensive educational dataset focused on diagram-driven course question generation. It contains:

• 25,798 questions
• 15,720 unique diagrams
• Coverage across 6 disciplines, 37 subjects, and 371 courses
• Average of 10.5 objects per diagram

Data Access

Due to certain data upload limitations on Hugging Face, some portions of the dataset may not be completely available through this platform.

To ensure full access to the complete DiagramQG dataset, we provide an additional download link via Quark Cloud Storage: https://pan.quark.cn/s/2bf45d6c982c?pwd=P64u

Please use this alternative link if you encounter any issues accessing the complete dataset files through Hugging Face.

Due to the ongoing peer review process of our research paper, we are currently releasing a subset of the DiagramQG dataset.

Dataset Structure

Discipline Areas

The dataset covers six main disciplines:

• Basic sciences (41.7%): Biology, Chemistry, Physics, Mathematics, Astronomy
• Earth and environmental sciences (14.9%): Geography, Geology, Climate Science, Environmental Science, Agriculture, Oceanography
• Healthcare and health (11.2%): Medicine, Physiology, Pharmacy, Public Health
• Business and economics (11.4%): Economics, Finance, Accounting, Management, Marketing
• Engineering and technology (10.2%): Computer Science, Electrical Engineering, Mechanical Engineering, Architecture, Design, Materials Science, Integrated Sciences
• Humanities and social sciences (10.6%): History, Psychology, Literature, Art Theory, Civics, Cognitive Science, Grammar, Music, Sociology, Writing

Hierarchical Organization

Data is organized hierarchically:

1. Discipline (e.g., Basic sciences)
2. Subject (e.g., Biology)
3. Course (e.g., Ecological interactions)

Data Collection Process

Phase 1: Initial Data Gathering

• Sources: Existing diagram-related datasets, Hugging Face, and Roboflow
• Raw dataset: 25,000+ diagrams and 60,000+ questions

Phase 2: Organization

• Classification into 6 disciplines and 37 subjects
• Mapping questions to 371 distinct courses

Phase 3: Annotation

• Experienced subject-specific annotators with relevant knowledge backgrounds annotate:
  • Input text constraints for each diagram-question pair
  • Course-specific contextual information
• Educational utility evaluation using 100-point scale

Phase 4: Quality Assurance

• Samples scoring below 70 points removed
• Only highest-scoring text constraints retained
• Final dataset: 25,798 validated samples

Dataset Analysis

Dataset Comparison

Dataset	Questions	Images	Objects/Image	Image Type	Text Constraint	Subjects
VQAv2.0	1.1M	20k	3.5	natural	answer	N/A
FVQA	5,826	2k	2.9	natural	answer	N/A
VQG-COCO	25,000	5k	3.3	natural	image caption	N/A
K-VQG	16,098	13K	2.7	natural	knowledge triple	N/A
TQA	26,260	3,455	7.2	diagram	-	10
ScienceQA	21,208	10,332	4.3	diagram & natural	-	12
MMMU	11,550	11,264	4.5	diagram & natural	-	30
DiagramQG	25,798	15,720	10.5	diagram	input, course	37

Unique Challenges

1. Domain-specific Knowledge Requirement

   • Unlike existing VQG datasets that rely on general common sense
   • Requires specialized knowledge across various disciplines
   • Models need to possess course-specific understanding

2. Long-tail Distribution Phenomenon

   • Course coverage ranges from abundant to limited samples
   • Challenges model generalization across well-represented and underrepresented courses
   • Reflects real-world educational resource allocation patterns

3. High Object Information Density

   • Diagrams contain significantly concentrated visual information (10.5 objects per diagram)
   • Complicates content interpretation and risks overlooking critical details
   • Demands enhanced visual processing capabilities

Novel Text Constraints

DiagramQG introduces a novel constraint mechanism that combines:

• Input constraint: Guides question generation around targeted diagram elements
• Course constraint: Ensures question relevance to specific educational contexts

This approach transcends traditional constraint formats like simple answers or image captions, enabling the generation of pedagogically meaningful questions that assess students' course understanding.

Data Splits

The dataset is split as follows:

• Training: 70% (17,880 questions, 11,817 diagrams)
• Validation: 5% (1,104 questions, 1,151 diagrams)
• Testing: 25% (5,565 questions, 6,767 diagrams)

Each course is represented in both validation and testing sets to ensure comprehensive evaluation across all educational domains.