Hugging Face fine-tuned the open-weights model Palette (20B) for generating presentations in the IBM Design Language
Hugging Face fine-tuned the 20B open-weights model Palette for creating presentations in the IBM Design Language. The model splits the work into a planning role and a coding role and, according to the company, took first place in two out of three domains of the PresentBench benchmark.
Hugging Face introduced Palette, a 20B open-weights model fine-tuned with a LoRA adapter for generating presentations in the IBM Design Language style. The model splits slide creation into two roles: the designer plans the entire presentation and writes a structured JSON brief (palette, typography, content, and layout for each slide), and the coder writes the pptxgenjs JavaScript code for an individual slide according to this brief. Because the context for each slide is fixed before the code is written, slide generation runs in parallel, which according to the company speeds up the coding phase roughly 6 to 7 times.
The training data was generated synthetically by a larger open teacher model — approximately 1000 presentations totaling about 10 000 slides. Each example also includes a rationale for decisions in the form of a counterfactual (why one variant was chosen over another, both in design and layout geometry). The generated slides went through automatic checks for shape overlaps and text overflow; faulty slides were corrected and became additional training examples. On the PresentBench benchmark, the model, according to the company, took first place in two out of three measured domains.
The motivation, according to the company, is that frontier models fall back to their own "home" style when generating presentations regardless of prompting, and that the content of corporate presentations sometimes cannot leave the company's own infrastructure for security reasons. A company's own small open-weights model solves these two problems. One of the key challenges during development was "content fidelity" — the model would invent facts not present in the input when filling out templates; the solution, according to the company, was not text instructions but sample reference presentations in the teacher model's prompt.
The article also includes a specific example of directives for adjusting the design, but the source text is cut off at this point. See the source article for details.
Why it matters
For companies with their own corporate design (colors, typography, layout) and sensitive presentation content, this approach offers an alternative to sending data to cloud frontier models: a smaller model can be run on the company's own infrastructure, trained on its own style, and, according to the reported results, still achieve comparable output quality. For developers and ML engineers, the described approach of splitting a complex generation task into a planning role and an execution role using a single LoRA adapter is of interest, along with synthetic data containing decision rationale and an automated verification mechanism for fixing errors — a technique applicable beyond presentation generation as well.
Two audiences, two different impacts
What this means
For individuals
Developers and ML engineers can take away a concrete approach for splitting complex output generation into a planning role (JSON brief) and a coding role (slide code) using a single LoRA adapter, including synthetic data with counterfactual reasoning and an automated verification harness for fixing faulty outputs.
For a business
Companies with their own design language and sensitive presentation content gain the option to run a smaller self-hosted model that, according to Hugging Face, achieves comparable quality to frontier models (first place in two out of three PresentBench domains), does not need to send data to the cloud, and thanks to parallel slide generation is faster in the coding phase.
ProductivityCheck the original
Event sources
only one source so far · 1 publisher, 0 independent. We count feeds from the same owner only once.