Dataset of chartQA (EMNLP 2025 Oral), designed to probe model sensitivity to common misleading visualization practices (cherry-picking, inappropriate scales, missing data, dual encoding, etc.). Each case includes a chart figure, underlying data, and multiple-choice QA with a correct option and a "misleader" option with explanation.

Directory layout: code/, data/, figures/, qa/. Paths follow <misleader_type>/<plot_type>/<case_name>.<ext>.

• Total cases: ~3,060
• Total files: ~12,240

dataset/
├── code/           # HTML visualization code (one file per case)
├── data/           # CSV data files
├── figures/        # JPEG chart images
└── qa/             # JSON question-answer files (question, options, correct, wrongDueToMisleader)

Each qa/*.json file typically contains:

• question: string
• options: list of four option strings
• correct: index of the correct answer
• wrongDueToMisleader: index of the option that is tempting from the chart but wrong given the data

The dataset covers many misleading visualization types, including (names may use underscores, e.g. MS_inappropriate_order):

• Cherry Picking, Exceeding The Canvas, Small Size
• MS Inappropriate Scale Functions / Scale Range / Inappropriate Order / Unconventional Scale Directions
• Dual Encoding, Missing Data, Inappropriate Aggregation
• Continuous Encoding For Categorical Data, Categorical Encoding For Continuous Data
• Misuse Of Cumulative Relationship, Data Visual Disproportion
• Concealed Uncertainty, Overplotting, Lack Of Legend, Lack Of Scales
• Misleading Annotations, Missing Normalization

Plot types include bar_chart, line_chart, area_chart, scatter_plot, pie_chart, stacked_bar_chart, choropleth_map, heatmap, etc.

• Training / evaluation: Use dataset/figures/ as images and dataset/qa/ as labels; align by case name (filename without extension).
• Reproducing charts: Use dataset/code/*.html with dataset/data/*.csv and a local HTTP server to re-render the same charts.

Please cite the original paper (https://aclanthology.org/2025.emnlp-main.695/) when using the data.