Palette 20B Open Model Tops Two PresentBench Deck Tests
Palette fine-tunes a 20B open model on IBM's design language, ranking second overall on PresentBench while a repair harness lifts scores by 3.5 points.
Summary
Detailed November 3, 2026, Palette fine-tunes a 20B open-weight model with one LoRA adapter to create self-hosted decks in the IBM Design Language, keeping confidential content and design control on company infrastructure. Its designer role turns plans into editable JSON briefs covering content, layout, palette, typography and rationale. Its coder role converts each slide specification into pptxgenjs JavaScript, with parallel generation making coding about 6 to 7 times faster. Because coder examples were five times longer, designer examples required fivefold oversampling. A large open teacher generated about 1,000 synthetic decks and 10,000 slides from controlled recipes, including directives across 12 categories and counterfactual reasoning. Corrected reference decks improved factual fidelity more reliably than added prompt rules.
Palette’s harness fixes known errors, renders slides, detects overlaps, clipping and overflow, then lets an editor model rewrite flagged code. Repairs are retained only when checks confirm improvement. Across 131 PresentBench decks, the harness added 3.5 points on average. PresentBench evaluates fundamentals, visual design, completeness, correctness and fidelity; a separate frontier model converted its source documents into plans. Palette ranked first in Economics and Education, third in Talk, and second across all three domains, scoring 59.7 against NotebookLM’s 60.8 while beating Manus 1.6 and five other systems. No benchmark yet measures adherence to a specific design language. Next steps include GRPO-style reinforcement learning using checker signals, extracting source figures with Docling, and reskinning decks from one or two existing .pptx templates without retraining.
Positives
- Palette ranked first in PresentBench’s Economics and Education domains using one 20B open-weight model.
- Parallel coder calls generate slides about 6 to 7 times faster than sequential coding.
- The repair harness improved scores by an average of 3.5 points across 131 PresentBench decks.
- About 1,000 synthetic decks and 10,000 slides taught Palette the IBM Design Language and counterfactual design reasoning.
- Repairs are accepted only when automated checks confirm improvement, preventing the harness from degrading the original slide.
Risks & concerns
- Palette scored 59.7 overall, remaining 1.1 points behind NotebookLM’s 60.8.
- Palette placed third in the narrative, speech-driven Talk domain, behind two closed systems.
- Sparse briefs can force the coder to invent content, and prompt rules alone did not reliably prevent factual failures.
- PresentBench required a separate frontier model to convert source documents into plans before Palette could generate decks.
- No available benchmark measures whether generated decks consistently follow a specified design language.