Full VisEval evaluation of direct LLM-to-Vega-Lite generation, evidence workbench, and an evidence-derived NL2VIS architecture.
-
Updated
Jul 30, 2026 - Python
Full VisEval evaluation of direct LLM-to-Vega-Lite generation, evidence workbench, and an evidence-derived NL2VIS architecture.
To associate your repository with the viseval topic, visit your repo's landing page and select "manage topics."