Palette: teaching a small model to build decks in your design language
IBM Research has introduced Palette, a system that fine-tunes a 20B open-weight model to build slide decks in IBM’s Design Language. The goal is a self-hostable model that keeps confidential content on internal systems.
What the article says
Palette splits the work into two roles served by one LoRA adapter. A “designer” role writes a structured JSON brief covering layout, content and palette. A “coder” role writes pptxgenjs code for each slide, running in parallel, which makes that stage about 6–7× faster. Users can steer output with plain-English directives in double brackets, such as “make this a timeline.”
The training data includes about 1,000 synthetic decks (roughly 10,000 slides). A check-and-repair harness renders each slide, flags overlaps and overflow, and keeps a fix only if it measurably improves the slide; it adds 3.5 points on average across 131 decks. On PresentBench, Palette ranks first in Economics and Education, third in Talk, and second overall, 1.1 points behind NotebookLM (59.7 vs. 60.8).
Read the official article AnnuIA profile
Summary written by AnnuIA in its own words from the official page of Hugging Face. The content and rights belong to the publisher: please read the original article.