During evaluation the model suffered
template collapse: many generations rolled back to the same 5-section landing template regardless of prompt. This is why a v3 pass on
real, scraped source code (
cipher-sft25-real-merged) was launched and is now the recommended checkpoint.
Synthetic SFT that distills from another LM converges on the teacher's mean aesthetic — very few outlier designs — and under-weights the long tail that makes Awwwards sites Awwwards sites. Symptoms:
The fix was real data: 578 Three.js examples, 148 Motion dev files, 63 GSAP gallery snippets, ≤998 aura shells → shipped as the cipher-real-v1-sft.jsonl (5.66 MB) dataset split.
CC-BY-NC-4.0. Gemma-4 terms apply.
