dantelok/multidomain-chart-qa

Multi-Domain Synthetic Chart QA (Verifiable Labels)

A synthetic chart question-answering dataset for training and evaluating Vision-Language Models (VLMs). Each example is a rendered chart image paired with a natural-language question and an answer that is computed from the underlying data, so every label is ground truth by construction — not a language model's guess.

Generated by the pipeline at Code-based-Synthetic-Multimodal-Data-Generation.

Reference build: ~12,600 QA pairs over ~1,800 charts, balanced across three domains (climate / e-commerce / housing) and four chart types (bar, pie, line, scatter), split 90/10 train/test. Rebuild it deterministically with build_hf_dataset.py --charts-per-domain 600 --seed 0.

Why this dataset

Most chart-QA data is scraped or has answers written by an LLM, which cannot guarantee correctness — exactly where it matters most (aggregations). Here the chart and the QA are produced from the same exact data slice: the chart is rendered deterministically with matplotlib, and the answer is computed with pandas. For a column cases = [100, 250, 400], "total" is 750 and "max" is 400, never a plausible-but-wrong value. This mirrors how DVQA, FigureQA, and PlotQA are constructed.

How it was built

For each source table the builder repeatedly:

1. picks a chart type (bar / pie / line / scatter) and columns,
2. aggregates an exact slice (e.g. total of a numeric column grouped by a category),
3. renders the chart deterministically, and
4. generates ground-truth QA over that same slice (sum, mean, max, min, range, argmax, argmin, count, compare, lookup).

The whole process is offline, seeded, and reproducible — no API calls. See build_hf_dataset.py in the source repo.

Data fields

Field	Type	Description
image	image	The rendered chart (PNG)
question	string	A natural-language question about the chart
answer	string	The ground-truth answer, computed from the data
chart_type	string	bar / pie / line / scatter
domain	string	climate / ecommerce / housing
source	string	Upstream dataset + license
op	string	Operation used to compute the answer (e.g. argmax)
columns	list[string]	Columns involved in the question

Splits

train / test (90 / 10), split with a fixed seed.

Baseline

A strong off-the-shelf VLM (Cohere command-a-vision-07-2025), zero-shot on a 120-example test sample: 79% on structural questions (which category / comparison / count), 22% on questions requiring an exact numeric value, 40% overall. The gap reflects a known weakness of current VLMs — reading chart structure is easier than reading or computing precise values — which is exactly what verifiable labels let you measure. Reproduce with evaluate_baseline.py in the source repo.

Source data & attribution

All sources are Creative Commons Attribution 4.0 (CC BY 4.0); this derived dataset is released under the same license.

• Climate — Our World in Data, CO₂ and Greenhouse Gas Emissions — CC BY 4.0 — https://github.com/owid/co2-data
• E-commerce — UCI Machine Learning Repository, Online Retail — CC BY 4.0 — https://archive.ics.uci.edu/dataset/352/online+retail
• Housing — Inside Airbnb, New York City listings — CC BY 4.0 — https://insideairbnb.com/get-the-data/

Limitations

• Read-off precision. Scatter/line answers (e.g. an exact maximum) may not be perfectly readable from the image alone; they are exact for the underlying slice.
• Template phrasing. Questions come from templates, so linguistic diversity is narrower than human-written questions (answers are what's guaranteed, not fluency).
• Aggregation choices. Bar/pie answers are over summed groups (as plotted); the op/columns metadata makes the computation explicit and auditable.
• Source snapshots. Values reflect the source files at build time.

Citation

@misc{synthetic_chart_qa,
  title  = {Multi-Domain Synthetic Chart QA (Verifiable Labels)},
  author = {Lok, Dante and Kaur, Avneet and Chagas Fernandes, Reuben},
  year   = {2025},
  howpublished = {\url{https://github.com/dantelok/Code-based-Synthetic-Multimodal-Data-Generation}}
}

License

CC BY 4.0. You must credit this dataset and the upstream sources listed above.