chart-reasoning-verified

Chart reasoning examples generated from an explicit latent representation. The data, the question and the answer are computed before the chart is drawn, so the image is a rendering of known ground truth rather than the source of it. No model was asked to label anything.

Each row carries both a rendered chart and a text serialisation of the same chart, so the set is usable for vision-language training and for text-only language model training without regenerating anything.

Contents

Split	Rows
train	6883
validation	298
test	319

7500 rows over 2428 distinct charts and 2428 independent latent datasets, with 3399 distinct question strings. Median completion length 19 words.

Splits are family-disjoint: one latent dataset produces several related rows, and every row of a family stays on the same side of the split boundary.

Task families

Family	Rows	Share
statistical_interpretation	2349	31.3%
ranking	2129	28.4%
comparison	1087	14.5%
quantitative_reasoning	1010	13.5%
trend	642	8.6%
abstention	283	3.8%

Chart types

Type	Rows	Share
bar	1921	25.6%
line	1695	22.6%
histogram	1650	22.0%
grouped_bar	1156	15.4%
scatter	770	10.3%
stacked_bar	308	4.1%

Rendering conditions

Style	Rows	Share
small_text	1101	14.7%
truncated_axis	1088	14.5%
legend_distance	1084	14.5%
dense_labels	1080	14.4%
clean	1079	14.4%
clutter	1048	14.0%
similar_colors	1020	13.6%

The abstention rows

283 rows ask a question the chart genuinely does not answer, and the correct completion declines to answer and says why.

These are produced, not authored. After each chart is rendered the axis limits are read back off the figure and compared against the values the question depends on. When a difficulty factor pushes those values outside the drawn area, the row becomes a refusal naming the visible range and the categories that fall outside it.

The rule is per operation, because a hidden point is not the same as a missing one. A value below the axis minimum is known to be smaller than everything drawn:

• argmax stays answerable. A hidden point cannot be the maximum.
• argmin is answerable only when exactly one point is hidden, since that one is then the minimum. Two hidden points cannot be ordered.
• comparison and trend stay answerable when the two points sit in different bands, because their order is known.
• difference and correlation need the magnitudes, so any hidden referenced point makes them unanswerable.

In the text serialisation, values outside the drawn range print as not shown, so the refusal is correct in both modalities.

Loading

from datasets import load_dataset, Image

ds = load_dataset("vinod-anbalagan/chart-reasoning-verified", data_files={
    "train": "train.jsonl",
    "validation": "validation.jsonl",
    "test": "test.jsonl",
})
ds = ds.cast_column("image", Image())

ds["train"][0]["image"]        # PIL image
ds["train"][0]["prompt"]       # same chart as text
ds["train"][0]["completion"]

Images are not duplicated: several rows reference the same chart, which is why this uses jsonl with an image path column rather than the imagefolder convention.

Fields

image, prompt (text serialisation plus the question), question, answer, completion, completion_mode, task_family, task_subtype, chart_type, chart_title, x_label, y_label, unit, visual_difficulty, difficulty_factors, operation, evidence, referenced_labels, labels, values, axis_y_min, axis_y_max, near_tie_delta, id, family_id, domain.

labels and values are the latent data the chart was drawn from, so any answer can be recomputed independently.

What is verified, and what is not

Verified. Every answer is recomputed from the latent data by an independent verifier before the row is written, covering all task operations. Every row is checked for answerability against the axis limits actually rendered. Splits are family-disjoint. No duplicate prompts.

Not verified. Whether the questions are representative of what people ask about real charts. Difficulty is controlled and recorded but not calibrated against human performance. The chart aesthetic is matplotlib defaults and does not cover the visual variety of charts in the wild.

Limitations

• Synthetic data with synthetic aesthetics. Generalisation to real-world charts is untested.
• Question phrasing comes from a template set, so surface variety is bounded even though the underlying data is not.
• Only English.

Generation

Produced by data-visualizations. Every row records its generation seed, chart family, style, rendered axis limits and value shape, so any example can be reproduced exactly.