manifesta/scientific-chart-qa-17k

Scientific Chart QA — 17,070 rows

A multimodal chart-interpretation dataset built around one idea: teaching a model when not to answer is as important as teaching it to answer.

One in seven questions is deliberately unanswerable from its figure, and the correct response is cannot be determined. Baseline VLMs overwhelmingly guess a plausible-looking number instead. That is the behaviour this set targets.

• 17,070 rows over 6,696 unique figures — 10,644 synthetic charts with exact ground truth, 6,426 real CC-licensed scientific figures
• 2,467 unanswerable questions where the answer is a refusal
• Decontaminated against ChartQA, CharXiv, ChartX, DVQA and PlotQA by perceptual hash, CLIP embedding, n-gram and MinHash
• 100% publishable licensing — zero rows from non-redistributable papers

Built for the Adaption AutoScientist Challenge (Data Visualization, Part 2).

Why this exists

Public chart-QA datasets are dominated by synthetic business charts, and every question in them has an answer. Real scientific figures are harder — GPT-4o scores around 47% on CharXiv against roughly 80% for humans — and real figures routinely don't contain what you ask for.

So this set does two things the common corpora don't: it uses genuine scientific figures at scale (forest plots, ROC curves, volcano plots, PCA scatters, survival curves, confusion matrices), and it explicitly rewards refusing to answer when the figure cannot support one.

What is in it

Task	Rows	Share	Answer
Summarization / takeaway	3,695	21.6%	free text, one sentence
Value reading	3,418	20.0%	numeric, ±5% relative
Comparison	2,988	17.5%	categorical
Unanswerable	2,467	14.5%	cannot be determined
Chart structure ID	1,931	11.3%	single lowercase term
Trend reasoning	1,603	9.4%	increasing / decreasing
Caption generation	968	5.7%	free text

58.2% of rows are exactly gradable (numeric, categorical or boolean — see the exactly_gradable column). The remainder are free-text summaries and captions, which need a judge. That column is there so you can split the set honestly rather than pretending free text is string-matchable.

22 chart types are represented, including ones the mainstream benchmarks barely touch: forest plots with confidence intervals, survival curves, volcano plots, small multiples, dual-axis charts, and log-log plots.

Ground truth, and how far to trust it

Synthetic rows (10,644). Rendered by a deterministic engine that emits a full ground-truth sidecar — the underlying data table plus feature annotations — for every chart. Every answer is recomputed from that data table rather than trusted from the generator's own metadata. Rows whose recomputed answer disagreed were dropped (0.3% drop rate).

These charts are deliberately hard. 30 difficulty tags are applied at render time, including near_equal_values, crossing_lines, log_log, dual_axis_scale_mismatch, colorblind_similar_palette, dense_legend, unlabeled_ticks and negative_values. A chart tagged unlabeled_ticks genuinely has no readable axis numbers — and its questions are unanswerable because of that, on purpose.

Real rows (6,426). Figures harvested from open-access literature with a hard licence gate. Answers are caption-grounded: the ground truth comes from the expert caption written by the paper's authors, and every row carries an evidence pointer to the caption span supporting it. No value-reading questions are generated against real figures, because we cannot recompute those from a data table — so we don't invent them.

Numeric tolerance is RELATIVE

Numeric answers are graded within ±5% of the true value, not by exact match. This follows the ChartQA relaxed-accuracy convention. A value read off a bar against a gridline is not accurate to four decimal places, and grading it as though it were would punish correct reading.

If you evaluate against this set, honour the numeric_tolerance column or you will substantially under-score any model.

Decontamination

Eight stages, run against a poison set built from the public chart benchmarks:

Stage	Method	Removed
1	pHash + dHash vs benchmark images (Hamming ≤ 6)	223 rows
2	CLIP embedding vs benchmark images (cosine ≥ 0.92)	197 rows
3	Intra-set image dedup	1,940 rows
4	Q&A n-gram + MinHash near-dup	70 rows
5–6	Verified-correctness gate, licence provenance gate	0 rows

Stage 2 matters: CLIP caught 197 rows that perceptual hashing missed — re-styled clones of benchmark charts that pHash reads as different images. A decontamination report with per-stage counts and audit samples ships in this repo as decontam_report.json.

Covered: ChartQA, CharXiv, ChartX, DVQA, PlotQA. Known gap: FigureQA, which has no clean mirror to hash against. Stated rather than hidden.

Licensing — all 17,070 rows are redistributable

Source	Rows	Licence
Synthetic (generated for this dataset)	10,644	Own work
Open-access literature	6,220	CC-BY-4.0
Open-access literature	164	CC-BY-SA-4.0
Open-access literature	42	CC0-1.0

The aggregate is CC-BY-SA-4.0 because of the CC-BY-SA slice. Filter license != "cc-by-sa-4.0" (164 rows) for an attribution-only subset.

An earlier build of this dataset contained 620 rows drawn from figures in papers under the default arXiv licence, which does not grant redistribution. Those were removed before publication, and the check is now a hard gate in the pipeline. Every figure carries its licence, licence URL and source URL in the scrape manifest.

Loading

from datasets import load_dataset

ds = load_dataset("manifesta/scientific-chart-qa-17k", split="train")
row = ds[0]
row["image"]       # PIL.Image — embedded in the parquet, no separate download
row["prompt"]      # question, ending in an <image1> reference token
row["completion"]  # answer

# the anti-hallucination slice
refusals = ds.filter(lambda r: r["task_type"] == "unanswerable")   # 2,467

# only what you can grade by exact match
gradable = ds.filter(lambda r: r["exactly_gradable"])              # 9,940

Fields

Field	Notes
prompt	Question, with a trailing <image1> token linking it to the image
image	The figure, embedded as bytes (HF Image() feature)
completion	The answer
instruction_raw	Question without the <image1> token
task_type / subtask	See the task table above
chart_type	22 values
answer_type	numeric · categorical · text · boolean
numeric_tolerance	Relative tolerance (0.05 = ±5%)
exactly_gradable	True if string/numeric-matchable; False for free text
verifiable	Whether the row passed the pipeline's verification
image_sha256	Integrity + decontamination key
source · source_domain · license	Provenance
difficulty	easy · medium · hard

All 17,070 images were decoded and every image_sha256 was confirmed against the embedded bytes before release.

Prompt phrasing

Questions are templated by task type, then routed through a deterministic paraphrase layer so the same question is not repeated verbatim across thousands of images. The most frequent single prompt covers 1.6% of rows (an earlier build had one prompt on 20.8%). 3,655 distinct prompt strings.

They remain templated. If you need free-form natural questions, this is not that dataset.

Limitations — read these

• Real-figure answers are caption-grounded, not pixel-grounded. The ground truth is the author's caption. If a caption is vague or overstates the figure, that propagates. There are no value-reading questions on real figures for exactly this reason.
• 62% of rows are synthetic. Realistic in style and genuinely hard, but not drawn from literature. Train on the mix, and be aware of it when reporting.
• The unanswerable share (14.5%) is far above the 5% originally designed. It is the point of the dataset, but a model trained on this may skew toward refusal. Downsample that slice if you want a more answer-eager model.
• chart_type is a single label per figure. A small number of scraped figures are multi-panel composites that get one label for the whole composite, and a few are low enough resolution that their axis text is not legible.
• Summarization and caption rows cannot be graded by exact match. Use exactly_gradable.
• FigureQA is not in the decontamination poison set (no clean mirror available).
• No human review at scale. Samples were inspected by hand; 17,070 rows were not.

Citation

@misc{scientific_chart_qa_17k,
  title  = {Scientific Chart QA: 17,070 evidence-grounded chart questions with a refusal slice},
  author = {Aivaras Navardauskas},
  year   = {2026},
  url    = {https://huggingface.co/datasets/manifesta/scientific-chart-qa-17k}
}

Built with Adaptive Data by Adaption. Real figures are © their respective authors under the Creative Commons licences recorded per row; please honour those terms and cite the source papers where you rely on a specific figure.