Home/Events/Palette: Teaching a 20B Model to Build Decks in IBM Design Language

Palette: Teaching a 20B Model to Build Decks in IBM Design Language

Developing
Confidence
70%
Impact: 60%
Updated 3m ago

Consensus Brief

Palette is a project that fine-tuned a 20 billion parameter model to create slide decks based on the IBM Design Language. It utilizes a two-role approach, separating design and coding tasks to improve the quality and speed of deck creation. The model was trained on synthetic data and has shown promising results on the PresentBench benchmark.

Sourced from
Primary: Hugging Face

What Changed Since Last Update

3m ago

New official source added: Hugging Face published an update on Tue, 03 No ("Palette: teaching a small model to build decks in your design language").

Claim Ledger

4 claims tracked across sources

Confirmed Fact

Palette ranks first on two of the three domains measured on PresentBench.

Confirmed Fact

The model was trained on around 1000 decks covering roughly 10,000 slides.

Confirmed Fact

The coding stage is about 6–7× faster due to parallel processing of slides.

Independent Finding

Small models often produce broken decks with overlapping shapes and text running off slides.

Role-Based Impact Analysis