Oriel — AI & technology worth noticing, every day.

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

Web page · https://huggingface.co/blog/ibm-research/palette-deck-building

First seen by Oriel on .

AI summary

Palette is a 20B open-weight model for generating slide decks in IBM Design Language, using a designer-coder workflow and a check-and-repair harness; it ranks first on two PresentBench domains.

  • A single LoRA adapter handles a designer that creates a structured JSON brief and a coder that writes pptxgenjs code for each slide; slide coding runs in parallel, about 6–7× faster.
  • Training used roughly 1,000 synthetic decks covering 10,000 slides, with counterfactual reasoning and checker-based repairs; coder examples were about five times longer, so designer examples were oversampled 5×.
  • The harness fixes known errors, checks rendered slides for overlap and overflow, and keeps rewrites only when checks show improvement; it added 3.5 points across 131 PresentBench decks.
  • Palette ranks first on Economics and Education, third on Talk, and second overall at 59.7, 1.1 points behind NotebookLM; PresentBench does not measure adherence to a specific design language.
  • The team identifies reference decks as more effective than prompt rules for improving teacher fidelity, and plans to explore reinforcement learning, figure-heavy decks, and template re-skinning.

#llm #slide generation #lora #presentation design

Generated by gpt-5.6-sol. Check the original for details.

Aliases

SourceExternal IDURL
Hugging Face Bloghttps://huggingface.co/blog/ibm-research/palette-deck-buildinghttps://huggingface.co/blog/ibm-research/palette-deck-building Read discussion