Palette: teaching a small model to build decks in your design language
Source
huggingface.co
Date
Why it matters
Demonstrates training a self-hostable small model on synthetic data in a specific design language, with a verifying harness, to compete with frontier systems. Useful pattern for brand-constrained generation with private content.
Key takeaways · AI-distilled
IBM Research's Palette splits deck building into two roles on one LoRAA cheap way to fine-tune a model by training a small add-on layer instead of changing all of the model's weights.Full definition → adapter: a designer writes a JSON brief with palette, type and per-slide layout, and a coder writes each slide's pptxgenjs code, in parallel, making coding about 6-7x faster.
Coder training examples ran about 5x longer than designer examples, so a naive data mix undertrained the designer until the team oversampled designer examples 5x.
Training traces are written as counterfactuals that weigh an alternative and say why it was rejected, such as setting a 62-character title at 36pt instead of 44pt so it fits on two lines, so the model learns the rule behind each choice.
Prompt rules did not stop the teacher model from inventing facts on slides. Fixing the reference decks in its prompt did, because the teacher copied those references far more than it followed instructions.
Palette's agent harnessThe scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.Full definition → keeps an editor model's slide rewrite only when the geometry checker confirms the slide improved, so repairs cannot make a slide worse; across 131 PresentBench decks it added 3.5 points on average.
Terms in this piece · Glossary
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
LoRA — A cheap way to fine-tune a model by training a small add-on layer instead of changing all of the model's weights.
