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

Why it matters

IBM's blog post introduces Palette, an open-weight model fine-tuned to build presentation decks that follow a company's own design language rather than the generic house style of frontier systems. The piece frames a practical problem: frontier models drift back to their defaults no matter how they are prompted, and confidential content may not be sent to external services at all. A self-hosted model solves ownership but small open models produce broken decks, since writing slide code requires unseen layout arithmetic and mixes editorial judgment with geometry in a single pass. Palette addresses this by splitting the work into two roles handled by one LoRA adapter: a designer that writes a structured JSON brief with design rationale, and a coder that writes the pptxgenjs code for each slide. Training data was synthesized by a large open teacher using recipes and counterfactual reasoning, with a checker that renders every deck, repairs broken slides, and turns each caught mistake into a better training example. The post also reports that reference decks steered the teacher more reliably than prompt rules, which fixed faithfulness failures.