VillanovaAI/multi_figureQA

multi_figureQA

Overview

Multi-FigureQA is visual question-answering (VQA) and reasoning dataset based on Microsoft's FigureQA dataset.

The original FigureQA dataset consists of synthetic, scientific-style figures (such as dot-line plots, line plots, horizontal and vertical bar graphs and pie charts) paired with source data and Q&A pairs.

While the original FigureQA dataset provides clean synthetic data for basic chart comprehension, its plots lack realistic labels, titles, and domain-specific context. As a result, the generated figures offer limited practical utility for contemporary visual reasoning tasks.

For more information about Microsoft's FigureQA VQA dataset check the realated paper on ArXiv.

---

To overcome the abstract nature of the original dataset, Multi-FigureQA introduces persona-driven figure generation. By using detailed persona descriptions, the pipeline generates domain-specific titles, descriptive labels, and realistic data distributions, significantly increasing the dataset's practical utility. Q&A pairs are then generated on the basis of the Persona and related figure.

In-prompt instructions ensure question ad topic diversity and strict answer format.

• Questions and answers rely exclusively on what is visibly displayed in the chart and described in the source code.
• The generation logic restricts questions when chart values are visually too close to differentiate, preventing unanswerable or ambiguous answers.
• To prevent over-representation of simple extreme values, questions are evenly distributed across four distinct categories: highest_lowest,pairwise_comparison, relative_ranking, presence_composition.
• Responses are structured into four constrained formats: yes_no,detailed_explanation, concise, label_only.

Dataset structure

{
    "image_index": int64, 
    "image": image, 
    "graph_type": string, 
    "source_code": string, 
    "persona": string, 
    "question_type": string
    "question": string,
    "answer_format": string
    "answer": string
}

Multilingual generation

The dataset contains approximately 200,000 samples, with 60% in English. The remaining 40% is distributed evenly among French, Spanish, German, and Italian (10% each). The final distribution is as follows:

Language	Subset Name	Percentage	Number of Examples
English	en	60%	119,920
Italian	it	10%	19,962
French	fr	10%	19,988
Spanish	es	10%	19,994
German	de	10%	19,948
Total		100%	199,812

Models

Figure and Q&A generation was achieved using the state-of-the-art, permissively licensed model Qwen/Qwen3.5-122B-A10B-FP8.

Licensing

The dataset is distributed under the ODC-BY-1.0 license. Users must adhere to the attribution and usage requirements outlined under this license.