vinod-anbalagan/gridline-chartqa

Adaption Charts P2 — Gold Chart-QA Dataset

A verified, quality-first chart question-answering dataset built for the Adaption Labs AutoScientist Challenge (Part 2, Data Visualization track). Two sources: a programmatically generated synthetic core (correct-by-construction) and a hand-authored hardset built from real public dashboards and reports.

At a glance

<!-- AUTOGEN:counts -->

• —1415 rows total — 1317 synthetic + 98 hardset <!-- /AUTOGEN:counts -->
• —7 chart types — bar, line, groupedbar, stackedbar, pie, donut, mixed
• —10 task types — every allowed task_type populated at target
• —English only, real-world business/finance/health/policy domains
• —Every row verified=true

Schema

Each row:

Composition

<!-- AUTOGEN:composition --> By source

By chart_type

By task_type

By difficulty

<!-- /AUTOGEN:composition -->

Perceptual difficulty

Most chart-QA datasets vary difficulty by arithmetic — more steps in the calculation. That is not what challenges a vision-language model. A capable VLM handles multi-step arithmetic easily once it has read the values; what it struggles with is reading the values in the first place.

358 rows (25.3%) are hard by perceptual construction. Eight mechanics:

Roughly 90 of these rows carry no value labels whatsoever — the answer must be read off gridlines. Answers remain correct-by-construction because values are generated as exact multiples of the tick step.

Every hard row's notes field names its mechanic, e.g. hard/truncated_axis; procedural gen (seed=777, idx=12).

Verification protocol

Every row is verified=true, but the mechanism differs by source:

Synthetic — correct-by-construction. Each chart is rendered from a seeded pseudorandom value distribution. The answer to every question is computed from those underlying values before the chart image is drawn. There is no visual estimation involved. The full generation pipeline lives in the companion repo alongside this dataset. Seeds are deterministic.

Hardset — hand-authored, human-reviewed. Every hardset row was authored one at a time from a real chart screenshot. Each row was reviewed row-by-row against the source image during authoring, with the answer recorded only after cross-checking what the chart actually shows. All questions were designed to be answerable from the image alone without external context.

Data sources (hardset only)

Hardset uses screenshots of publicly available charts from official statistical / research bodies. Each row's notes column carries the source institution and original screenshot filename.

Attribution notes. WHO source materials are typically licensed under CC-BY-NC-SA-3.0 IGO. This dataset uses WHO chart screenshots for the transformative purpose of vision-language model training, with full attribution preserved in each row's notes. Downstream users concerned about commercial use should filter rows where notes begin with hardset; WHO; and treat them separately.

Intended use

Fine-tuning multimodal chart-QA models. Short-answer chart understanding benchmarks. Ablation studies on chart-type or task-type coverage.

Not intended for downstream tasks that require gold-standard OCR fidelity or exact numerical extraction from dense financial tables — this dataset targets reasoning about charts, not exact digit extraction.

Known limitations

• —English only.
• —Synthetic aesthetic is uniform. All synthetic rows use the same matplotlib renderer with a consistent style. Real-world visual diversity comes entirely from the hardset rows.
• —Hardset skews toward `max_min` and `lookup_value`. Real dashboards naturally support these tasks; the balance step trims synthetic max_min heavily to compensate, but hardset structural bias remains.
• —`chart_type=mixed` is exclusively hardset — synthetic doesn't generate multi-panel dashboards. Mixed charts are visually harder for VLMs and represent a natural difficulty axis.
• —No held-out split in this file. An external evaluation slice (ChartQAPro-derived) lives outside this dataset by design.

Citation

If you use this dataset, please cite:

@misc{anbalagan2026adaptioncharts,
  title  = {Adaption Charts P2: A Small Gold Chart-QA Dataset},
  author = {Anbalagan, Vinod},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/vinod-anbalagan/adaption-charts-p2-gold}}
}

Please also credit the upstream chart sources listed in the "Data sources" section above when relevant.

License

This dataset is released under CC-BY-4.0. See attribution notes above for source-specific considerations.