Nutrient benchmark
Chart Parsing Leaderboard
Chart image in, structured JSON out: title, chart type, data table, encoding, axes, series, legend, and data labels. Models see the same frozen charts and the same prompt, and are ranked by mean per-chart F1 within 3% tolerance.
Submit or reproduce a result
Generate one raw output per benchmark chart with the shipped prompt, score it with the reference scorer, and submit the resulting JSON for review. The benchmark set is frozen and will not change silently.