ChartCraft: the chart-editing dataset of ChartReformer (ICDAR 2024)

This is the official dataset release of ChartReformer from 2024. The paper was published at ICDAR 2024 (arXiv: March 2024). The data here is the original dataset from the 2024 paper, unchanged. It was uploaded to Hugging Face in October 2026 only to make it easier to download. Before that, it was available only as a Google Drive archive linked from GitHub. This is not a new dataset, a new version, or new work.

• 📄 Paper: ChartReformer: Natural Language-Driven Chart Image Editing (ICDAR 2024) · arXiv:2403.00209
• 💻 Code: github.com/pengyu965/ChartReformer
• 🤖 Model: Pengyu965/ChartReformer
• 📚 Dataset: Pengyu965/ChartCraft (this repo)

Pengyu Yan, Mahesh Bhosale, Jay Lal, Bikhyat Adhikari, David Doermann. University at Buffalo.

Overview

ChartReformer is the first work to systematically study chart image editing: changing a chart image from a natural-language instruction, without access to its original data or plotting code. To make this task learnable and measurable, the paper defines a taxonomy of chart edits and builds ChartCraft. To our knowledge, ChartCraft was the first publicly available dataset for chart editing.

Each sample is an edit pair:

• the original chart image and its full structured description: the underlying data table, title, axis titles, and all visual attributes;
• one or more natural-language edit instructions (up to seven paraphrases);
• the edited chart image and its structured description after the edit.

Because both sides of every pair are fully specified, ChartCraft supports chart editing (image + instruction → edited chart) and chart de-rendering (image → data table + visual attributes).

Edit taxonomy

Category	edit_type values
Style	common/font, common/font_size, specific/colors, specific/hatches (bar), specific/linestyles and specific/markers (line)
Layout	grid, legend
Format	conversion to line, grouped vertical bar, stacked vertical bar, grouped horizontal bar, stacked horizontal bar
Data-centric	data_filter/range_based, data_filter/series_based, data_point_change/add_data_point, data_point_change/remove_data_point, data_series_change/add_data_series, data_series_change/remove_data_series

Five chart types are covered: line, grouped vertical bar, stacked vertical bar, grouped horizontal bar, and stacked horizontal bar.

Splits

Split	Edit pairs	Style	Layout	Format	Data-centric
train	88,800	16,800	8,000	16,000	48,000
validation	11,100	2,100	1,000	2,000	6,000
test	5,550	1,050	500	1,000	3,000
test_small	555	105	50	100	300

test is the 5.5K test set used for the main results in the paper. test_small is a separate held-out set, 10% the size of test, with the same edit distribution. It is meant for quick evaluation, for example of slow baselines; the paper compares against ChartLlama on a test set of this size. All splits are disjoint.

Every chart type appears in every split (each split is about 21% line charts and about 20% each bar type). Data-centric edits are sampled more often because they are the hardest.

Data fields

Field	Type	Description
id	string	Sample ID (the original image filename without .png)
chart_type	string	Chart type of the original chart
edit_category	string	style_edits, layout_edits, format_edits, or data_edits
edit_type	list[string]	Full edit path, e.g. ["line", "data_edits", "data_point_change", "add_data_point"]
editing_prompts	list[string]	Natural-language edit instructions (paraphrases of the same edit)
original_image	image	The chart before editing (PNG)
edited_image	image	The chart after editing (PNG)
underlying_data	struct	data_table, chart_title, x_axis_title, y_axis_title of the original chart
visual_attributes	string (JSON)	Plotting parameters of the original chart: global_properties plus line_properties or bar_properties
edited_underlying_data	struct	Same as underlying_data, for the edited chart
edited_visual_attributes	string (JSON)	Same as visual_attributes, for the edited chart

In data_table, rows are separated by <0x0A> and cells by | (the DePlot convention). Visual attributes are stored as JSON strings because line and bar charts have different attribute sets. Use json.loads to read them.

Usage

import json
from datasets import load_dataset

ds = load_dataset("Pengyu965/ChartCraft", split="test")
ex = ds[0]

ex["original_image"]          # PIL image of the chart before the edit
ex["editing_prompts"][0]      # e.g. "Introduce grid lines to the plot"
ex["edited_image"]            # PIL image of the chart after the edit

# Chart-editing target used to train ChartReformer: data first, then visual attributes
target = json.dumps({**ex["edited_underlying_data"], **json.loads(ex["edited_visual_attributes"])})

The training split is about 8.5 GB. Use streaming=True to iterate without downloading everything first.

Export to the original folder layout

The GitHub training code reads the original release format: images/, edited_images/, and summary.json in each split folder. This snippet rebuilds that layout, byte-for-byte, from the Hugging Face version:

import json, os
from datasets import load_dataset, Image

def export_split(split, out_dir):
    ds = load_dataset("Pengyu965/ChartCraft", split=split)
    ds = ds.cast_column("original_image", Image(decode=False)).cast_column("edited_image", Image(decode=False))
    os.makedirs(os.path.join(out_dir, "images"), exist_ok=True)
    os.makedirs(os.path.join(out_dir, "edited_images"), exist_ok=True)
    summary = {}
    for i, ex in enumerate(ds, start=1):
        name = ex["id"] + ".png"
        for col, folder in (("original_image", "images"), ("edited_image", "edited_images")):
            with open(os.path.join(out_dir, folder, name), "wb") as f:
                f.write(ex[col]["bytes"])
        summary[str(i)] = {
            "original_image": f"images/{name}",
            "editing_prompts": ex["editing_prompts"],
            "edit_type": ex["edit_type"],
            "edited_image": f"edited_images/{name}",
            "visual_attributes": json.loads(ex["visual_attributes"]),
            "edited_visual_attributes": json.loads(ex["edited_visual_attributes"]),
            "underlying_data": ex["underlying_data"],
            "edited_underlying_data": ex["edited_underlying_data"],
        }
    with open(os.path.join(out_dir, "summary.json"), "w") as f:
        json.dump(summary, f)

for split, folder in [("train", "train"), ("validation", "val"), ("test", "test"), ("test_small", "test_small")]:
    export_split(split, f"data/{folder}")

How the data was made

• Data tables: The tables come from AdobeSynth-19, released with the ICPR 2020 CHART-Infographics competition. Its underlying data is derived from real-world sources such as the World Bank's World Development Indicators and Gender Statistics.
• Rendering: A custom Matplotlib tool renders each chart. Its visual attributes (colors, line styles, markers, hatches, fonts and font sizes, legend location and columns, grid, tick labels) are sampled at random from a large parameter pool, so the charts look very different from each other.
• Edit pairs:
  • Style and layout edits change one plotting parameter.
  • Format edits keep the data table and the color of each series, and re-plot the chart as a different chart type.
  • Data-centric edits change the data table and keep the visual style consistent.
• Instructions: Each pair comes with one or more instructions. Several paraphrases of a base instruction were generated so the data covers different ways of phrasing the same edit.

Limitations

• The charts are synthetic and rendered with Matplotlib. Real-world charts made with other tools can look different.
• Only line charts and grouped and stacked bar charts are included. Pie charts, scatter plots, and others are not.
• Instructions are template-based and in English. Each describes a single, concrete edit.

Citation

If you use ChartCraft or ChartReformer, please cite:

@inproceedings{yan2024chartreformer,
  title={ChartReformer: Natural Language-Driven Chart Image Editing},
  author={Yan, Pengyu and Bhosale, Mahesh and Lal, Jay and Adhikari, Bikhyat and Doermann, David},
  booktitle={International Conference on Document Analysis and Recognition},
  pages={453--469},
  year={2024},
  organization={Springer}
}

License

Released under the MIT License, the same as the GitHub repository. The data tables are derived from AdobeSynth-19 (ICPR 2020 CHART-Infographics). Please also follow the terms of that source.