Palette: teaching a small model to build decks in your design language

| Source: Hugging Face Blog

Tags: IBM Research, Palette, PresentBench, IBM Design Language, synthetic data, fine-tuning, Hugging Face

IBM Research's Palette fine-tunes a 20B open model on synthetic IBM Design Language data to build slide decks in a house style, ranking first on two of three PresentBench domains.

Details

IBM Research argues that frontier models drift back to their own house style when generating slide decks, and confidential content sometimes can't leave your systems, which points to a self-hosted model trained on your own design language. Small open models usually produce broken decks: invisible titles, overflowing text and overlapping shapes, because they lay out slides in code without seeing them and must make editorial and geometric decisions at once. Palette fine-tunes a 20B model with a planner-designer and a slide-coder split, trained on synthetic data following the IBM Design Language and wrapped in a checking harness. IBM says it ranks first on two of the three PresentBench domains it measured.