Palette: teaching a small model to build decks in your design language
Palette demonstrates that a 20B open‑weight model fine‑tuned with LoRA on synthetic IBM Design Language data can generate slide decks that outperform frontier models on the PresentBench benchmark, ranking first on two of three measured domains. The system splits deck creation into a “designer” phase that produces a structured JSON brief and a “coder” phase that emits pptxgenjs JavaScript for each slide, enabling parallel generation and a 6–7× speedup. This approach lets enterprises keep models and data on‑prem while achieving high‑quality, brand‑consistent decks.
⚡ Key Takeaways
- 20B model fine‑tuned with LoTS (LoRA) ranks first on two of three PresentBench domains.
- Architecture separates responsibilities: a designer model writes a JSON brief, a coder model writes pptxgenjs code.
- Parallel coder execution yields a 6–7× reduction in deck generation latency.
- Integration uses the pptxgenjs JavaScript library for slide rendering.
- Requires synthetic training data aligned with the IBM Design Language and a LoRA adapter; full fine‑tuning is not used.
Engineers can now host a modest‑size model in‑house to produce brand‑consistent slide decks without exposing confidential content, while still matching or beating large proprietary models on benchmark quality.
✅ Practical Steps
- Generate synthetic training data that follows your organization’s design language (e.g., IBM Design Language) and train a LoRA adapter on a 20B base model.
- Deploy the trained adapter to produce a JSON brief (designer) and then feed each slide spec to the coder model, which outputs pptxgenjs code.
- Run the coder model in parallel across slides to achieve the reported 6–7× speedup.
Want the full story? Read the original article.
Read on Hugging Face Blog ↗