Cipher SFT 2.5 — Synthetic (v2, retired) 🦑⚠️
Superseded by
cipher-sft25-real-merged. Archived for reproducibility.
This is the v2 Awwwards SFT checkpoint of the Cipher series, trained on a synthetic dataset of 288 Awwwards-style triples distilled from a larger instruction-tuned model.
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.
- 🧠 Base:
Auroraventures/cipher-simpo-merged - 📚 Dataset:
awwwards-stage25-sft.jsonl(288 synthetic records, 1.9 MB) - ⚠️ Status: Retired — use v3 instead for production
- 📦 GGUF:
cipher-sft25-merged-Q4_K_M-GGUF
Why retired?
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:
- Identical hero layouts across divergent prompts
- Repeated color tokens (
#9bf,#a4f) regardless of brief - Repeated section ordering: hero → about → services → work → contact
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.
License
CC-BY-NC-4.0. Gemma-4 terms apply.
Built with 🦑 by Aurora Ventures.
- Downloads last month
- 840