SCAFFOLD

SCAFFOLD: A Large-Scale Structured Dataset of Computer Science Research Figures with Diagram QA and Chain-of-Thought Reasoning Traces is a large-scale multimodal reasoning dataset designed for training and evaluating Vision-Language Models (VLMs) on scientific figure understanding and visual reasoning.

The dataset is constructed from figures extracted from publicly available arXiv research papers and contains 157,387 question-answer pairs covering diverse scientific reasoning tasks. Each question is paired with a scientific figure and is accompanied by metadata such as reasoning type, difficulty, confidence, and an optional chain-of-thought reasoning trace.

SCAFFOLD is intended for research in Scientific Visual Question Answering (VQA), Multimodal Reasoning, Scientific Document Understanding, and Vision-Language Instruction Tuning.

Note: This dataset is released for research purposes only. Any unwanted use can lead to the violation of the intended terms of use and may result in legal action or revocation of access rights.

Motivation

Existing multimodal reasoning datasets primarily focus on synthetic charts, general-purpose images, or document understanding tasks. Comparatively few datasets target real scientific figures extracted directly from research publications while simultaneously providing rich reasoning annotations suitable for instruction tuning.

SCAFFOLD bridges this gap by providing a large-scale collection of real scientific figures paired with diverse reasoning questions and chain-of-thought annotations. The dataset covers a broad spectrum of scientific visual content, including charts, plots, algorithms, architecture diagrams, flowcharts, conceptual illustrations, and experimental visualizations.

---

Key Contributions

• 29,887 real scientific figures extracted from publicly available arXiv papers
• 157,387 multimodal question-answer pairs
• 3,058 research papers
• Figure-level reasoning annotations
• Multiple visual reasoning categories
• Difficulty and confidence annotations
• Optional chain-of-thought reasoning traces
• Instruction-tuning ready chat format
• Paper-level metadata for every source document

---

Dataset Statistics

Metric	Count
Research Papers	3,058
Scientific Figures	29,887
Total QA Pairs	157,387
Training Instances	120,225
Validation Instances	36,797

---

Supported Tasks

SCAFFOLD supports research in:

• Vision-Language Instruction Tuning
• Scientific Visual Question Answering (VQA)
• Scientific Figure Understanding
• Scientific Diagram Understanding
• Chart and Plot Reasoning
• Multimodal Chain-of-Thought Reasoning
• Scientific Document Understanding
• Vision-Language Benchmarking

---

Question Categories

The dataset contains diverse reasoning tasks including:

• Boolean Reasoning
• Label Reading
• Counting
• Component Identification
• Comparative Reasoning
• Relational Reasoning
• Structural Reasoning
• Causal Reasoning
• Object Identification

---

Dataset Structure

SCAFFOLD/
│
├── README.md
├── LICENSE
├── records.jsonl
├── train.jsonl
├── validation.jsonl
├── pdf_metadata.jsonl
└── images.zip

---

Dataset Files

File	Description
images.zip	Archive containing all extracted scientific figures organized by paper ID.
records.jsonl	Complete dataset with every QA instance, metadata, and reasoning traces.
train.jsonl	Training split formatted for multimodal instruction tuning.
validation.jsonl	Validation split using the same format as the training data.
pdf_metadata.jsonl	Per-paper metadata including extraction statistics and QA distributions.

---

Image Organization

Images are grouped by their source paper.

Example:

images/
└── 1112.1768/
    ├── 1112.1768_p08_fig00.png
    ├── 1112.1768_p09_fig01.png
    └── ...

Each filename follows the convention

{paper_id}_p{page}_fig{figure_number}.png

Example:

1112.1768_p08_fig00.png

which corresponds to the first extracted figure on page 8 of paper 1112.1768.

---

Dataset Schema

records.jsonl

Each record contains metadata describing a single visual reasoning instance.

Field	Description
figure_id	Unique identifier of the QA instance
paper_id	Source arXiv paper identifier
figure_number	Figure number within the paper
caption	Figure caption (if available)
ref_sentence	Reference sentence extracted from the paper
question	Question associated with the figure
answer	Ground-truth answer
question_type	Visual reasoning category
difficulty	Difficulty level (easy, medium, hard)
confidence	Annotation confidence
has_cot	Indicates whether a reasoning trace is included
cot_trace	Chain-of-thought reasoning trace
image_path	Relative path to the corresponding figure

---

pdf_metadata.jsonl

Contains document-level metadata including:

• Number of extracted figures
• Number of generated QA pairs
• Question-type distribution
• Difficulty distribution
• Confidence distribution
• Figure extraction statistics

---

train.jsonl / validation.jsonl

Training and validation sets are stored in multimodal chat format suitable for instruction tuning.

{
  "messages": [
    {
      "role": "user",
      "content": [
        {
          "type": "image",
          "image": "images/1112.1768/1112.1768_p08_fig00.png"
        },
        {
          "type": "text",
          "text": "What algorithm is presented in the figure?"
        }
      ]
    },
    {
      "role": "assistant",
      "content": [
        {
          "type": "text",
          "text": ""
        }
      ]
    }
  ]
}

---

Data Collection Pipeline

SCARLET was generated through a fully automated data generation pipeline consisting of:

1. Selection of publicly available arXiv papers licensed for redistribution.
2. Figure extraction from research paper PDFs.
3. Figure organization and metadata generation.
4. Prompt construction using figure metadata.
5. Automatic generation of question-answer pairs.
6. Generation of chain-of-thought reasoning traces.
7. Metadata aggregation and dataset splitting.

The complete data generation pipeline, preprocessing scripts, and utilities used to construct SCARLET are provided in the Code section of this dataset.

Additionally, scripts for downloading the original arXiv PDFs corresponding to the released dataset are also available in the Code section.

---

Intended Uses

SCAFFOLD is suitable for:

• Vision-Language Model pretraining
• Multimodal instruction tuning
• Scientific Visual Question Answering
• Scientific figure understanding
• Scientific chart understanding
• Scientific diagram reasoning
• Multimodal reasoning research
• Chain-of-thought supervised fine-tuning
• Benchmarking multimodal reasoning systems

---

Limitations

SCAFFOLD is automatically generated and should be used with the following considerations:

• Question-answer pairs are generated automatically and may occasionally contain inaccuracies.
• Chain-of-thought traces are machine-generated and should not be considered human-authored explanations.
• Some figures contain dense scientific notation or small text that may be difficult for current VLMs.
• Figure captions and reference sentences are unavailable for a subset of figures due to limitations in PDF extraction.
• The dataset is derived from arXiv publications and may not fully represent scientific literature from other publishers or disciplines.

---

Biases

Users should be aware of several potential sources of bias:

• The dataset is dominated by English-language scientific publications.
• Paper selection is limited to publicly available arXiv papers.
• Certain research domains (e.g., computer science, machine learning, and mathematics) are more heavily represented than others.
• Question-answer annotations reflect the capabilities and biases of the language model used during automatic generation.

---

Citation

If you use SCAFFOLD in your research, please cite:

@misc{raut2026scaffoldlargescalestructureddataset,
      title={SCAFFOLD: A Large-Scale Structured Dataset of Computer Science Research Figures with Diagram QA and Chain-of-Thought Reasoning Traces}, 
      author={Ranjit Raut and Aarav Subedi and Sagun Rai and Sudan Jha},
      year={2026},
      eprint={2609.00018},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2609.00018}, 
}

---

License

The annotations, metadata, dataset organization, and generated reasoning traces are released under the (CC BY NC 4.0)** license.

Scientific figures are extracted from publicly available arXiv papers. Copyright for each figure remains with its respective authors and is governed by the license of the corresponding source paper.

Each dataset instance contains the source paper_id, enabling users to trace every figure back to its original publication.

---

Acknowledgements

SCAFFOLD was created by automatically extracting scientific figures from publicly available arXiv research papers and generating multimodal reasoning annotations using large language models.

The authors gratefully acknowledge the arXiv repository and the original authors of the source publications for making their research openly accessible.