Ask for a chart from a table. Both models generate code, both are executed, and you see the charts that actually came out.
Model · Dataset (ChartForge) · trained on 112,000 rows where every code sample was executed before it shipped.
Research demo built for the Adaption AutoScientist Challenge (Data Visualization). The GPU backend scales to zero when idle, so the first request may take up to a minute. Generated code runs in an isolated container, matplotlib only, with a hard timeout, so a non-terminating program is reported rather than left running. The models are trained on synthetic tables, so real-world spreadsheets with missing or messy values may behave less predictably than the examples here.