01 / The short version
What happened
Palette is a system that fine-tunes a 20B open-weight model to generate presentation decks using a specific design language, addressing the tendency of frontier models to revert to their default styles. It splits the design process into two roles: a designer for editorial choices and a coder for layout geometry, using a LoRA adapter. Trained on synthetic data generated by a teacher model and validated by a rendering harness, it ranks first on two of three PresentBench domains. The system allows users to apply specific directives for design variations and ensures data confidentiality by running on local infrastructure.
See the exact references02 / Key takeaways
What you need to know
- 01
Palette uses a two-role system (designer and coder) to improve small model performance in deck generation.
- 02
Synthetic training data is generated with counterfactual reasoning and validated through a rendering harness to catch layout errors.
- 03
User directives wrapped in [[ ]] allow fine-grained control over specific design elements like charts and colors.
Keep in perspective
What to watch for
The source text does not specify the exact release date, pricing, or availability of Palette, and performance claims are attributed to the publisher without independent verification.
Go to the source
Exact references
These are the original pages used for this brief. Publisher claims are not independent evaluations.
01Primary source · Hugging FaceRead the original announcementhttps://huggingface.co/blog/ibm-research/palette-deck-buildingAI-generated from the linked source. It can miss context; verify consequential details in the original. How the radar works