🔍 Read the full analysis: How Palette Teaches A Small Model To Build Decks In Your Design Language 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 says it fine-tuned Palette, a 20-billion-parameter open-weight model, to create slide decks in IBM Design Language using separate planning and coding roles. The company reports top rankings in two of three PresentBench domains and faster slide coding, but has not supplied scores, evaluation details or a public release timetable in the source material.
As detailed in the original analysis, Hugging Face says it fine-tuned Palette, a 20-billion-parameter open-weight model, to build presentations in IBM Design Language by splitting deck planning from slide-code generation. The company reports that Palette ranked first in two of three PresentBench domains it evaluated, though the supplied account gives no scores or testing details to assess the result independently.
Palette uses two roles through a single LoRA adapter, a fine-tuning method. First, a designer role turns a prompt into a structured JSON brief covering the deck’s content, palette, typography, layouts and design rationale. A coder role then converts each slide specification into pptxgenjs JavaScript that defines the slide layout. Hugging Face says the separation lets slide coding run in parallel, making that stage about six to seven times faster than sequential coding.
The training material was synthesized with a larger open model. Hugging Face says the lack of a public corpus connecting finished decks to their plans, design decisions and code led the team to generate about 1,000 decks, or roughly 10,000 slides. The examples varied in genre, length, layout and input style, and included explanations of why a design choice was selected over an alternative. The post says coder examples were about five times longer than designer examples, so designer examples were oversampled fivefold during training.
The team also describes a quality-checking loop: generated decks were rendered and inspected for overlapping elements and text overflow. Broken slides were repaired, with the associated reasoning revised into further training examples. Hugging Face says the process also addressed a faithfulness problem: when instructions lacked detail, the model could invent content. In the team’s account, examples drawn from reference decks guided the teacher model more reliably than adding further prompt rules.
A Workflow for Branded Presentations
Palette’s approach addresses two practical barriers to automated presentations: general-purpose models may not follow a company’s visual system, and some organizations may not want confidential material processed by an external service. Hugging Face presents a fine-tuned, open-weight model as a way for teams to adapt the output to their design language and potentially run the system on their own infrastructure.
Those are project motivations, not demonstrated deployment outcomes. The source does not report that Palette has been deployed inside an organization, show how it handles sensitive information, or establish that its performance transfers beyond IBM Design Language. The reported benchmark ranking suggests a promising result on some measured tasks, but without domain names, scores, competitors or evaluation conditions, it cannot establish broad superiority.
The structured brief is also a practical design choice: it makes deck-wide decisions explicit before code is written, and the plan can be reviewed or edited by a person or another agent. That could make errors easier to catch than in a workflow where a model produces a finished deck in one pass. Whether that translates into fewer revisions or better presentations in regular workplace use remains unreported.
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
Samsung Odyssey G4 25-inch FHD Gaming Monitor
Jackery Explorer 300 Portable Power Station, 292Wh
As an affiliate, we earn on qualifying purchases.
Why Planning Comes Before Slide Code
Presentation generation combines editorial decisions with geometric ones. A system must decide what information belongs on each slide, then place text and visual elements at readable sizes and positions. Hugging Face says small models can struggle when they write slide code without seeing a rendered result: text may overflow, elements may overlap, and titles may become difficult to read.
Palette separates those jobs. Its designer role makes deck-wide content and style choices, while its coder role handles individual slide layouts. The project’s rendering and repair loop feeds visible layout problems back into training examples. This account describes the team’s method; it does not provide independent tests showing how often the checks catch errors or how much the revisions improve final decks.
The post frames the method as adaptable to a user’s design language, but the reported work centers on IBM Design Language. It does not include results for other design systems or explain what customization would require. The model is described as open-weight, but the supplied material does not name its base model or state a release license.
“The designer handles the editorial judgment, and the coder handles the geometry.”
— Hugging Face, describing Palette’s approach
Benchmark and Release Details Missing
The PresentBench claim cannot be independently judged from the supplied account: it does not identify the three domains, scores, competing systems or evaluation conditions. The reported six-to-sevenfold speed improvement applies to slide coding compared with sequential coding, according to Hugging Face; the post does not provide timing measurements or clarify whether the figure includes the full deck-generation process.
Other open questions include how Palette performs on complex data, long presentations, strict factual requirements and design systems beyond IBM’s. The source also does not show independent testing, deployment results or the frequency with which the model adds unsupported information. It does not specify the model’s base weights, license, safeguards for private infrastructure or a date for releasing weights, code or evaluation results. Until those details are available, the benchmark ranking and speed figure remain claims by the project’s authors.
Public Weights and Wider Testing
The supplied post does not give a release date or say whether Palette’s weights, code or detailed evaluation results will become publicly available. If those materials are released, developers and organizations could test the workflow on their own decks and compare its output across other design systems.
Useful follow-up evidence would include PresentBench scores and full evaluation conditions, plus tests of factual accuracy, layout failures and performance on longer or data-heavy presentations. Results from deployment on private infrastructure would also help clarify the operational and privacy claims. For now, the next step is not specified by Hugging Face in the source material.
Key Questions
What is Palette?
Palette is a 20B open-weight model that Hugging Face says it fine-tuned to create slide decks in IBM Design Language.
How does Palette generate a deck?
It uses a designer role to produce a structured JSON brief, then a coder role to write pptxgenjs JavaScript for individual slide layouts. Hugging Face says this split allows coding to run in parallel.
How strong were Palette’s benchmark results?
Hugging Face reports that Palette ranked first in two of three PresentBench domains it measured. The source account does not provide domain names, scores, competitors or test conditions, so the result cannot be compared quantitatively here.
Can organizations download and run Palette now?
The supplied material does not say whether the model weights or code are publicly available, identify a license, or give a release date. It describes open weights and on-premises use as relevant motivations, not confirmed deployment details.
Primary source: Hugging Face · via ThorstenMeyerAI.com
