Hugging Face Develops Palette To Create Custom Decks

Hugging Face reported that current frontier models can generate good slide decks, but they often lack customization and may not meet specific design requirements. According to Hugging Face, these models have their own house style, which can be difficult to change.

Small open models, on the other hand, are not effective at generating decks due to their limited capabilities. They often produce broken decks with titles that vanish into the background, text running off the slide, and shapes stacked on top of each other. Hugging Face's Palette aims to address this issue by fine-tuning a small open-weight model to build decks in a user's design language.

Palette's approach involves training a model to build decks in a way that mimics human designers and coders. The model is split into two roles: a designer that writes a structured JSON brief, and a coder that writes the JavaScript code to draw each slide. This approach allows for more control over the design process and enables the model to make editorial judgments and geometric decisions separately.

The training data for Palette is synthesized using a large open teacher, which generates decks and their reasoning. Each deck starts from a recipe that sets its genre, length, layouts, and any user directives. The model learns to cut and condense content, and the data includes reasoning behind each design decision. The teacher also writes counterfactuals to explain why certain design choices were made.

The model is trained with a focus on faithfulness, ensuring that the output is accurate and reliable. Reference decks are used to fix issues with the model's output, and the teacher is trained to copy these reference decks rather than following rules. This approach enables the model to learn the intent behind the design choices rather than just following instructions.

Palette also allows users to steer the deck in plain English using directives, which are short instructions written directly in the plan. These directives can be used to specify design choices, such as using a chart instead of a table or changing the color scheme. The model is trained to follow these directives and can learn the intent behind them.

The development of Palette is significant because it enables users to create custom and confidential slide decks using a small open-weight model. This approach has the potential to improve the efficiency and effectiveness of deck creation, and it allows users to own the model and its output. According to Hugging Face, Palette's approach can compete with frontier systems while keeping ownership of the result.