🔍 Read the full analysis: Inside Palette: Training A Small Model To Make Design-Matched Decks on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Hugging Face describes Palette, a fine-tuned 20-billion-parameter open-weight model for generating slide decks in IBM Design Language. The company reports first-place rankings in two of three PresentBench domains and says parallel slide coding made that stage about six to seven times faster, but it provides no scores or evaluation details in the supplied account.
Hugging Face says it fine-tuned Palette, a 20-billion-parameter open-weight model, to generate slide decks following IBM Design Language, as described in the original analysis. The company reports that Palette ranked first in two of three PresentBench domains it measured, but the supplied account gives no scores, domain names or comparison details, leaving the result difficult to assess independently.
Palette uses one LoRA adapter for two distinct roles. A “designer” role converts a user’s input into a structured JSON brief that describes the deck’s content, palette, typography, slide layouts and design rationale. A “coder” role uses specifications for individual slides to write JavaScript for pptxgenjs, the library used to create the presentation file. Hugging Face says separating planning from implementation lets slide coding run in parallel and makes that stage about six to seven times faster than running it sequentially.
The training examples were generated using a larger open model, according to the project account. Hugging Face says no public dataset contains the plans, design choices and code behind finished decks, so it created about 1,000 decks covering roughly 10,000 slides. The examples varied in genre, length, layout and input style, and included explanations of why a particular design choice was made instead of an alternative.
The team says it rendered generated decks and checked them for overlapping elements and text overflow. It repaired broken slides and added revised reasoning to the training examples. The account also describes a faithfulness issue: when a brief did not specify enough, Palette could fill gaps with invented content. Hugging Face says examples drawn from reference decks steered the teacher model more reliably than adding further prompt rules. The supplied material does not quantify how often that problem occurred or how often the checks caught layout defects.
A Structured Route to Branded Decks
Palette targets two practical concerns with general-purpose presentation tools: their slides may not follow an organization’s design system, and sending confidential material to an outside service may be unsuitable. A model that can be adapted to a company’s visual language and run on its own infrastructure could give teams more control over both the deck’s appearance and where inputs are processed. Those are motivations for the project, not demonstrated deployment outcomes; the account does not show Palette operating in a production environment or establish how it handles sensitive information.
The workflow’s separation of editorial planning and slide geometry is also a specific design choice. It creates a structured brief before code is written, which could make the deck’s content and visual decisions easier for a person or another agent to review. Hugging Face presents this split as a way to address errors that occur when a model writes slide code without first planning the deck. Whether the approach produces more accurate or usable presentations across varied tasks remains to be tested against detailed, comparable results.
The reported benchmark ranking is a signal of performance on some measured tasks, not evidence that Palette outperforms larger or frontier models overall. Without scores, baselines and evaluation conditions, readers cannot determine the size of the result or how it relates to other systems.
Products Worth Considering
Bose SoundLink Flex Bluetooth Speaker (2nd Gen) – Portable Outdoor Speaker with Hi‑Fi Audio, Waterproof and Dustproof, USB‑C, Up to 12 Hours
EcoFlow Delta 3 Portable Power Station – 1024 Wh, 1800 W
Samsung Odyssey G4 25-inch FHD Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How Palette Splits Deck Work
Building a presentation requires both deciding what belongs on each slide and translating those decisions into positions, font sizes and shapes. According to Hugging Face, smaller models can produce layout problems when they generate code without reviewing a rendered slide: text may overflow, elements may overlap, or titles may be difficult to read. Palette’s deck-wide JSON brief is intended to establish content and visual choices before the coder produces slide files.
Hugging Face says coder examples in its training data were about five times longer than designer examples. To avoid undertraining the designer role, the team oversampled those examples fivefold. It used LoRA, a method for adapting a model with a smaller set of trainable parameters, which the account says made experiments less expensive while adding the presentation-generation task to the base model.
The project’s account focuses on IBM Design Language, while saying the approach can be trained for a user’s own design language. It does not report results for other systems. The available material also does not identify the base model used, give a public model release or license, or provide independent tests of generated decks.
“The designer handles the editorial judgment, and the coder handles the geometry.”
— Hugging Face, describing Palette’s workflow
Benchmark and Release Details Missing
The supplied account does not identify the three PresentBench domains, provide scores, name the competing systems or describe the evaluation conditions. It is therefore not possible to judge the margin behind the two reported first-place rankings or compare Palette quantitatively with other models. The reported six-to-sevenfold speed improvement is also attributed to the project; the account does not give timings, hardware details or a broader measure of end-to-end deck creation time.
Other open questions include how Palette performs on complex data, strict factual tasks, longer presentations and design systems beyond IBM’s. The project acknowledges that incomplete specifications can lead to invented content, but gives no frequency or error-rate figures. It also does not describe operational safeguards for use on private infrastructure. The account does not announce when model weights, code or fuller evaluation results will be released, or whether they will be publicly accessible.
Independent Tests Would Clarify Performance
The next useful evidence would be published benchmark details: domain names, scores, baselines, competing systems and evaluation procedures. A public release of model weights and code, if one is planned, would let developers reproduce the workflow and test it on their own presentations. The supplied account does not give a release date or confirm that such a release is forthcoming.
Further testing could show whether the approach works across different design languages and more demanding decks, and how reliably the checking process detects layout failures or unsupported claims. Until those details are available, Palette’s ranking and speed figures should be understood as claims reported by the project team, rather than independently established results.
Key Questions
What is Palette?
Palette is a 20-billion-parameter open-weight model that Hugging Face says it fine-tuned to generate slide decks in IBM Design Language. Its workflow uses a designer role to plan the deck and a coder role to generate slide code.
How does Palette create a presentation?
The designer role creates a structured JSON brief with content and design choices. The coder role turns slide specifications into JavaScript for pptxgenjs. Hugging Face says the slide-coding stage can run in parallel.
How did Palette perform on PresentBench?
Hugging Face reports that Palette ranked first in two of three measured domains. The supplied account does not provide domain names, scores, comparison systems or evaluation conditions, so the result cannot be assessed in detail.
Is Palette publicly available?
The supplied material does not state whether Palette’s model weights or code have been released, provide a license or give a release date.
Does Palette guarantee accurate, private presentations?
No such guarantee is established in the account. Hugging Face says the model could invent content when a specification lacked detail, and describes checks for layout problems. It provides no quantified accuracy results or operational evidence about privacy protections.
Primary source: Hugging Face · via ThorstenMeyerAI.com
