Gemma4-Glamour-31B

Gemma4-Glamour-31B

A web-design generator fine-tuned from google/gemma-4-31B-it. Give it a natural-language brief; it returns a self-contained front-end as three fenced blocks — html, css, js.

Part of the Glamour project (Schneewolf Labs): an arena + synthetic-data pipeline for web-design generation and critique.

Training

• Method: QLoRA SFT (r=64, α=32), merged to full weights.
• Data: ~445 brief→implementation pairs. Trained on the Opus-authored subset only and evaluated on the held-out non-Opus subset (gemini, gpt, grok, glm, kimi, deepseek, haiku) — eval loss fell 0.67→0.42 across the run, improving even on designs authored by other models.
• Checkpoint: this is the phase-1 generator, taken at an early-stop checkpoint (~step 56/84) on the eval plateau. Format adherence is still developing; treat as a v1.

Usage

from openai import OpenAI            # or any chat client
c = OpenAI(base_url="http://localhost:8000/v1", api_key="x")
c.chat.completions.create(
    model="glamour-31b",
    messages=[{"role": "user", "content": "Design a neobrutalist pricing page."}],
)

Output is html / css / js fenced blocks. A Q4_K_M GGUF runs in llama.cpp at ~10 tok/s on a single GB10.

License

Apache-2.0 (inherited from Gemma 4).