nutrientdocs/chart-parsing-benchmark

Chart Parsing Benchmark

Chart image in, structured JSON out. A fixed, test-only benchmark for turning a chart image into its title, chart type, data table, encoding, axes, series, legend, and data labels. It holds 2500 charts and is the set behind the leaderboard. Every entry sees the same charts and the same prompt, and the reference scorer ships in this repository, so any result here can be reproduced.

• 🎯 Model: nutrientdocs/chart-parsing-vlm
• 🧪 Try it: nutrientdocs/chart-parsing-demo
• 🏆 Leaderboard: nutrientdocs/chart-parsing-leaderboard

chart type	charts
line	552
bar	463
pie	249
scatter	235
area	224
box	171
bubble	164
heatmap	100
histogram	54
treemap	54
combo	51
funnel	49
radar	46
violin	33
stem	25
step	15
candlestick	15

Task and prompt

Every entry on the leaderboard uses the same system prompt and instruction, shipped with the scorer. The model returns one JSON object. The target column holds the reference JSON as a string.

Metrics

The ranking metric is the mean per-chart cell F1 at 3% tolerance (mean_f1_tol3): each chart's data cells are scored for precision, recall, and F1, and the per-chart F1 is averaged so every chart counts once, whatever its table size. Cells are never pooled across charts.

A cell matches when its row identity (the first column plus every text column) and its column agree and the value is close enough. Row order never matters. Tables with a single value column, or whose rows all repeat one value, match on row identity alone. Three tolerance levels are reported side by side:

metric	a value counts as correct when it is
mean_f1	strict: relative 1e-3
mean_f1_tol3	exact (relative 1e-3) when the chart prints data labels, 3% in percentage points for share charts (pie, donut, 100%-stacked, treemap), otherwise 3% of the chart's value range. Histogram bin-edge columns are not scored; histograms are scored on the per-bin frequency
mean_f1_tol5	exact (relative 1e-3) when the chart prints data labels, 5% in percentage points for share charts (pie, donut, 100%-stacked, treemap), otherwise 5% of the chart's value range. Histogram bin-edge columns are not scored; histograms are scored on the per-bin frequency

The 3% level is the ranking metric because it matches what a careful reader can recover from the pixels; the strict level exposes models that copy printed labels well, and the 5% level shows how much is lost to small estimation error. Also reported: table shape match, per-field accuracy (chart type, title, axis labels, scales, legend, data labels, column types and roles), and set F1 for encoding and series.

Scoring

python score.py --benchmark-repo nutrientdocs/chart-parsing-benchmark \
  --predictions preds.jsonl --out result.json --name my-model

preds.jsonl has one {"id": ..., "text": ...} line per chart, where text is the raw model output. The scorer is self-contained and needs only datasets. Submit the resulting JSON with the model ID, license, and runtime details for review. Entries on the leaderboard carry one of three labels:

label	meaning
commercial	Nutrient model, weights under a commercial Nutrient license
provider	cloud model reached through its vendor API
open	open-weight model served locally with the benchmark prompt

Schema

from datasets import load_dataset
benchmark = load_dataset("nutrientdocs/chart-parsing-benchmark", split="test")

Field	Meaning
id	Stable chart identifier
image	RGB chart image
chart_type, title	Flat copies of target fields for filtering
category	Source label from ChartGen
data_source	Always render in this set
target	Reference JSON as a string

License and attribution

Charts derive from SD122025/ChartGen-200K, released under CC BY 4.0. This benchmark keeps the same license. The Nutrient Chart Parsing model evaluated on it is commercial; to evaluate or deploy it, contact Nutrient.

About the author

This project is maintained and funded by Nutrient - The deterministic document infrastructure enterprises run their highest-stakes workflows on: replayable output, clear exceptions, and full audit trails on the messy, regulated documents where AI alone breaks.