AttackViz

AttackViz is a chart-image dataset for studying correct and misleading data visualizations. It was introduced in the paper ChartAttack: Testing the Vulnerability of LLMs to Malicious Prompting in Chart Generation.

Each example contains a rendered chart image, metadata about the chart and question type, the expected gold answer, a binary label indicating whether the chart is correct or misleading, a misleading-visualization category, and serialized chart annotations.

Dataset Structure

The dataset contains 18,247 examples across three splits:

Split	Examples
train	6,059
val	6,146
test	6,042

Fields

• image: chart image.
• chart_type: chart family. Values include v_bar, h_bar, and line.
• question: natural-language question associated with the chart.
• question_class: question category. Values include compound, comparison, min_max, data_retrieval, and arithmetic.
• gold: list of expected answer strings.
• correct_or_misleading: binary label where 0 means correct and 1 means misleading.
• misleader: misleading visualization type, or none for correct charts.
• annotations: list of serialized JSON strings containing the underlying chart specification and rendering metadata.

Chart Types

Type	Examples
v_bar	7,414
h_bar	7,348
line	3,485

Question Classes

Class	Examples
compound	4,581
comparison	4,258
min_max	4,172
data_retrieval	3,989
arithmetic	1,247

Label Distribution

Label	Meaning	Examples
0	correct	6,373
1	misleading	11,874

Misleading Types

Type	Examples
none	6,373
inappropriate_use_of_stacked	2,689
misrepresentation	1,694
inverted_axis	1,674
inappropriate_axis_range	1,370
3d	1,276
inappropriate_use_of_log_scale	942
inappropriate_use_of_line	548
truncated_axis	498
ineffective_color_scheme	498
inappropriate_item_order	460
dual_axis	225

Loading

from datasets import load_dataset

dataset = load_dataset("jgermanmx/AttackViz")
print(dataset)
print(dataset["train"][0])

Intended Use

This dataset can be used to evaluate or train models for chart understanding, misleading visualization detection, visual question answering over charts, and analysis of how design choices affect interpretation.

Citation

If you use AttackViz, please cite:

@misc{ortizbarajas2026chartattack,
  title = {ChartAttack: Testing the Vulnerability of LLMs to Malicious Prompting in Chart Generation},
  author = {Jesus-German Ortiz-Barajas and Jonathan Tonglet and Vivek Gupta and Iryna Gurevych},
  year = {2026},
  eprint = {2601.12983},
  archivePrefix = {arXiv},
  primaryClass = {cs.CL},
  doi = {10.48550/arXiv.2601.12983},
  url = {https://arxiv.org/abs/2601.12983}
}

Notes

The annotations field stores JSON as strings. Parse individual entries with json.loads when structured access to chart specifications is needed.