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Superseded by cipher-sft25-real-merged.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
Safetensors
Model size
33B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Auroraventures/cipher-sft25-merged

Finetuned
(1)
this model
Quantizations
1 model

Dataset used to train Auroraventures/cipher-sft25-merged