cipher-sft-merged

Cipher SFT (v1 base) 🦑

First taste of the Kraken. The creative-web code generator that started it all.

Cipher-SFT-Merged is the v1 SFT checkpoint of the Cipher series — a 31 B Gemma-4 fine-tune that writes complete, single-file HTML documents in the Awwwards idiom. This checkpoint provides the foundation that all later Cipher stages (SimPO, SFT 2.5, GRPO) are built on.

• 🧠 Base: unsloth/gemma-4-31b-it-unsloth-bnb-4bit
• 🔬 Fine-tune: Unsloth QLoRA SFT on early Awwwards-style code pairs
• 🎨 Specializes in: hero sections, scroll experiences, glassmorphism, WebGL, Three.js, GSAP
• ⚙️ Merged weights: BF16, ready for HF Inference Endpoints or llama.cpp via the GGUF sibling

🔄 Upgrade path: For the best Cipher experience, see cipher-sft25-real-merged — the v3 checkpoint trained on real scraped Awwwards code.

---

Quickstart

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tok = AutoTokenizer.from_pretrained("Auroraventures/cipher-sft-merged")
model = AutoModelForCausalLM.from_pretrained(
    "Auroraventures/cipher-sft-merged",
    torch_dtype=torch.bfloat16, device_map="auto",
)

messages = [
    {"role": "system", "content": "You are Cipher, the Code Kraken. Emit complete single-file HTML."},
    {"role": "user", "content": "Build a portfolio landing with Lenis smooth scroll and GSAP reveals."},
]
prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(prompt, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(**inputs, max_new_tokens=4096, temperature=0.7)[0]))

Deploy as HF Inference Endpoint

Click Deploy → Inference Endpoints (dedicated) at the top of this page. Recommended: AWS us-east-1, Nvidia L4 ×1, min replicas 0. Then call via /v1/chat/completions (TGI OpenAI-compatible route).

Cipher Pipeline Map

License

CC-BY-NC-4.0. Base model governed by Gemma Terms of Use.

---

Built with 🦑 by Aurora Ventures. Part of the Kin runtime.