WB-ChartExtract

WB-ChartExtract is a benchmark for chart-to-table extraction: recovering the underlying numerical table from a chart image. It accompanies the paper Self-Ensembling Vision-Language Models for Chart Data Extraction.

It is built from World Bank Open Data time series (52 indicators, 218 countries, 65 years) and is designed to be more challenging than existing benchmarks: charts contain on average 7× more datapoints than ChartQA, print no value labels (removing the OCR shortcut), and span 4 chart types × 4 rendering libraries with heavy stylistic variation.

• Code: https://github.com/tberkane/vlm-ensemble-chart

Contents

The dataset contains 1,000 charts:

png/            # 1,000 chart images (1.png ... 1000.png)
tables/         # 1,000 ground-truth tables (1.csv ... 1000.csv)
metadata.json   # per-image metadata

• png/<id>.png — the chart image.
• tables/<id>.csv — the clean ground-truth table. The first column is the year (index); remaining columns are one series per country. Missing values are nan.
• metadata.json — a dict keyed by image filename with fields: chart_type (line / area / grouped_bar / stacked_bar), library (matplotlib / seaborn / plotly / bokeh), countries, series_name, num_years, and subsampled.

Chart type and library are assigned uniformly at random (≈62–63 charts per combination). Within each chart, font, color palette, grid, line style, markers, transparency, and figure size are randomized.

Usage

from huggingface_hub import snapshot_download
path = snapshot_download("tberkane/WB-ChartExtract", repo_type="dataset")

License

Released under CC BY 4.0. Derived from World Bank Open Data, which is also CC BY 4.0.

Citation

@inproceedings{berkane2026selfensembling,
  title     = {Self-Ensembling Vision-Language Models for Chart Data Extraction},
  author    = {Berkane, Thomas and Wang, Qianyi and Majumder, Maimuna S.},
  year      = {2026}
}