Datasets:
license: cc-by-4.0
task_categories:
- image-to-text
tags:
- charts
- chart-understanding
- data-extraction
- benchmark
- synthetic
size_categories:
- n<1K
configs:
- config_name: default
data_files: metadata.jsonl
ChartParse-Bench
58 synthetic charts whose exact values were written before the pixels existed.
Every gate in chart extraction measures self-consistency: redraw the extraction, check it lands on the same ink, call it verified. That passes a series traced confidently off the wrong axis. Because each chart here ships the array it was drawn from, the question stops being is this answer plausible and becomes is this answer right.
Contents
| family | charts | series | what it isolates |
|---|---|---|---|
| A — line | 18 | 62 | identity through crossings, resampling honesty, axis binding |
| B — scatter | 11 | 22 | recall under overplotting, size/shape channels, outliers |
| C — area | 9 | 29 | stacking, computed baselines, composited fills |
| D — compound | 12 | 43 | non-Cartesian frames, linked panels, mixed geometries |
| M — market | 8 | 38 | correlated series that all turn in the same session |
| total | 58 | 194 | 77,929 ground-truth points |
1–12 series per chart. 8 panels on log axes (3 log-y, 3 log-x, 2 log-log).
Each chart isolates one named failure, declared in its spec before it is drawn — A01 "colour-based tracing that swaps identity at a crossing", A14 "gap honesty", C07 "hatching is annotation", D10 "a curvilinear coordinate system", M06 "accumulated identity error: no single crossing is hard, there are simply a great many".
Files
images/<ID>.png the figure, as an extractor receives it
truth/<ID>.truth.json the exact values used to draw it, plus axis_scales and
`printed` flags on labels that actually appear
spec/<ID>.spec.json what the chart tests, its difficulty level, and for the
M set the measured crossing schedule
metadata.jsonl one row per chart
index.json the same rows as a single array
Suggested metric
score = series_recall x coverage x accuracy x mark_recall
Kept as a product, not an average: recovering one of eight series perfectly must not read as a good parse. Three properties matter more than the formula:
- Pair series by shape, not by name (Hungarian assignment), so a system cannot win on labelling.
- Error relative to each series' own range, and compared in log10 space on a log axis — otherwise a four-decade residual is dominated entirely by the top decade.
- Compare nothing where the extractor declared a gap. A declared absence is neither credited as coverage nor bridged and charged as error. Saying "I could not see this" has to be allowed to be the right answer.
Use RMS rather than median: a median hid a log-axis magnitude error at 12% where RMS showed 22%.
Known limitations
- C05 is drawn with no numeric y ticks, so its density values are not recoverable from the image at all. It tests shape only.
- A14 and C03 store values across a span wider than the figure can show (tails decaying to ~4e-6 of the peak). A coverage measure taken against the stored array rather than against visible ink will understate a correct parse on those two.
- D10 is a Smith chart stored in reflection coefficient. A system answering in normalised impedance is not wrong; the conversion is arithmetic an extractor should not be expected to perform.
- Synthetic throughout. These charts are matplotlib output, so they carry none of the compression artefacts, rotated scans, or hand-drawn annotation of a real document. They measure numeric correctness, not robustness to bad inputs.