Everyday plots from CSV tables¶

This example combines six common plots from machine learning, engineering and business: training curves, benchmark points, a sensor response, residuals, grouped revenue bars and a composition heatmap. All inputs are original simulated data, supplied as small typed CSV files. SVG/PDF authoring uses only Inklet's core dependencies; PNG export adds the render extra.

Full-size figure · Editable SVG · PDF · Executable recipe · Data and source hashes · LaTeX caption

Reproduce it¶

For a smaller step-by-step example, use From CSV to a figure. Run these commands from a checkout of this documentation revision, which includes the updated source recipe:

python -m pip install -e '.[render]'
python examples/general_plots.py

Outputs in out/general-plots/ include SVG/PDF/PNG, an HTML review, data.json, caption.txt and caption.tex. The input files are in examples/data/general-plots/; nothing is downloaded. In a LaTeX figure environment, place the PDF and then use \input{caption.tex}. Figure titles, panel descriptions and methods stay in the external caption.

The example reads explicit numeric types with inklet.read_csv, shares model colors between a and b, and uses live dataset references for plots. Revenue shares are derived from the same columns used for the grouped bars. Updating those columns updates both panels when the document is compiled again. The data guide explains source hashes and parsing rules.

Plot appearance changes¶

Each row pairs related views: loss with model comparison, response with residuals, and revenue with its composition. Blue, teal and rust consistently identify series. The benchmark labels name the points directly, while the response separates light sample markers from the darker model line. A reference line marks zero residual, and histogram separators make bin boundaries visible.

Axes use the theme's hairline weight, with light grey horizontal guides. The loss and response legends sit in empty corners of their data areas; the revenue legend sits in one row above the bars, which fill the corners. The matrix uses the default sequential colour ramp. Plot margins are shared so data areas align across the grid. Use axes and text for font and stroke controls, and plot layout for legends and insets. These guides describe the current behavior.

Compare the previous styling¶

The earlier styling comparison preserves the original six-panel layout with heavier axes and one-column legends. To compare those two styling choices in the current layout, run:

python examples/general_plots.py --legacy-look --output out/general-plots/before

This uses the current renderer and arrangement; it does not reproduce the archived image pixel for pixel. The main figure is 183 mm wide, the preset's double-column width. Rebuild at the required physical size instead of shrinking the finished image.

Read the plots¶

• a–b: Simulated loss curves and benchmark points share model names and colors. No actual model training, held-out evaluation or inference timing was performed.
• c–d: A simulated first-order response has Gaussian noise with standard deviation 0.03 and seed 2026. Residuals are supplied in the CSV and histogram bins are fixed at width 0.02. Inklet does not fit the response model.
• e–f: Simulated revenue in thousands of euros, and each product's percentage of the total for that quarter. The color scale spans 0–100%; the table provides the unnormalized inputs.

All data, processing, figure artwork and captions are MIT, Mark Marosi.