IBM trained a 20B model to build decks in a design language

IBM Research taught a small open-weight AI model to build presentation decks in the IBM Design Language. The post is on Hugging Face. It names the method, the training data and the measured gain.

The model is 20B parameters and runs on IBM's own hardware. IBM calls the system Palette. The team wrote the code that renders the slides in a library called PptxGenJS. The deck follows the template the user provides.

How the two roles split the job

Palette splits the work between a designer and a coder. One LoRA adapter does both. The designer reads the brief and writes a structured JSON file. The file names the palette, the typography, the layout of each slide, the content, and the reason for the choice. The coder then writes the PptxGenJS code for each slide.

A user can push the design toward a goal in plain English. The brief carries a directive in brackets like [[make this a timeline]] or [[this is for the board]]. IBM lists six kinds of directive: layout, tone, audience, order, leaving out and emphasis.

The checker, and the gain it adds

Every deck passes through a check-and-repair step. The step renders each slide. It looks for overlapping text, clipped content, broken colour codes or text that runs off the slide. An editor model rewrites the broken slide. IBM only keeps the repair if it measurably helps the slide.

The step adds 3.5 points per deck on average on IBM's own benchmark, PresentBench. The parallel coder runs the slides at once. IBM reports it as 6 to 7 times faster than the sequential version.

Where the model wins, and where it loses

IBM measured Palette on PresentBench across five criteria: fundamentals, visual design, completeness, correctness and fidelity. The team reports the model as the best in class on economics and education decks. On a talk-style deck, Palette scored 59.7 against NotebookLM's 60.8. That puts it in second place by 1.1 points on the three-domain mean. The post says the model is strongest on structured, numeric content. It is weaker on narrative, speech-driven decks.

Three findings a fine-tuning team can borrow

IBM states three findings in plain words. Put the reason behind a design choice in the training example, beside the choice itself. Fix the reference decks before writing a prompt rule, because the teacher model follows the examples more than the rules. Build the checker first, so the data is trustworthy and the output is safe.

The training set itself is small. About 1,000 decks and 10,000 slides. A larger teacher model wrote them to a recipe for genre, length, layout and directive. Each training example carries a reason: why the choice was made, and which other choice the designer thought about and rejected.

PresentBench is open on Hugging Face. The post does not say when the Palette model itself will be released.

Source

Hugging Face, with the PresentBench benchmark.