Palette: teaching a small model to build decks in your design language
Reported by Hugging Face1 source ↓
Hugging Face says it fine-tuned a 20B open-weight model to create slide decks in the IBM Design Language. The model ranked first in two of three PresentBench domains the team measured.
Key points
- The model uses separate designer and coder roles to plan deck content and render slides
- Training data covered about 1,000 decks and 10,000 slides
- Hugging Face says parallel slide-coding runs made that stage about 6–7× faster
- The post does not specify when or how Palette will be made available
Details
Hugging Face reports that Palette was fine-tuned with LoRA, using synthetic training data based on the IBM Design Language. It separates deck creation into a designer role that plans slide content and a coder role that places elements and draws each slide. The team says the model ranked first in two of the three PresentBench domains it measured.
The training set was generated with a large open model and contained about 1,000 decks, or roughly 10,000 slides. Hugging Face says each deck was rendered and checked for overlaps and text overflow; broken slides were repaired and added as training examples. The post also says the coding stage runs about 6–7 times faster because slides can be coded in parallel after the plan is set.
Written from coverage as of
Images
- Hugging Face
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh