Palette trains a 20B model to create decks in a custom design language

Palette is an open-weight system designed to generate presentation decks in a user-defined design language while running on private infrastructure. Its developers fine-tuned a 20B model with LoRA on synthetic data based on the IBM Design Language. The system separates the work into two roles: a designer creates a structured JSON brief covering content, layout and design rationale, while a coder produces pptxgenjs JavaScript for each slide. The slides can be generated in parallel, making the coding stage about 6–7 times faster.

The training set contained around 1,000 decks and roughly 10,000 slides. Examples included counterfactual reasoning and were rendered and checked for overlap and overflow; defective slides were repaired and added back to the data. Users can also add directives such as [[make this a timeline]] to control presentation choices.

A harness checks and repairs generated slides, retaining a rewrite only when measurements show improvement. On 131 decks, it raised PresentBench scores by an average of 3.5 points. Palette ranked first in Economics and Education, third in Talk, and second overall across the three domains, scoring 59.7 versus NotebookLM’s 60.8.