ML-builder: Tool to recreate charts with prompt-friendly inputs (ml-builder.vercel.app)

ML-builder is a new utility that reconstructs complex AI/ML figures from simple, prompt-friendly inputs—turning textual descriptions of plots into reproducible charts. The demo shows it can recreate a wide range of visualizations common in model papers and benchmarks: scaling-law line plots (e.g., perplexity or bits-per-word vs tokens/compute), mean pass-rate curves on coding tasks, grouped and stacked bar charts (exam percentiles, multi-model accuracy across languages and tasks), heatmaps (pairwise win rates, attack/task matrices), scatter/regression plots, error-bar comparisons, and even radial or multi-panel grids. It handles important plot mechanics such as logarithmic axes, error bars, annotations, subplots, and different grouping/stacking schemes. For the AI/ML community this matters because it automates a tedious but crucial step—faithfully reproducing and inspecting reported visual evidence across papers, benchmarks, and model audits—improving reproducibility, faster figure iteration, and easier verification of claims (e.g., inverse-scaling trends, adversarial robustness, cross-lingual accuracy). Technically, ML-builder maps high-level semantic descriptions (metrics, models, axes, groupings) into concrete plotting primitives and layout rules, enabling consistent replication of complex visuals like multi-model comparisons, compression ratios per language, and evaluation subsampling analyses. That makes it a practical tool for researchers, reviewers, and data journalists who need trustworthy, editable reproductions of published charts.