aayushale00/romax-card-designer-local

Romax Card Designer Local

Romax Card Designer Local is a fine-tuned SmolLM2 135M specialist model for generating compact membership-card design intent as structured JSON.

It is designed for the Romax Pass AI application. The model chooses design intent such as mood, orientation, brand asset treatment, field choice, and color usage. The application owns rendering, QR safety, brand locks, field limits, PNG export, and wallet-ready JSON.

Files

File	Purpose
card-designer-q4_k_m.gguf	Quantized local model for llama.cpp
local-concept.schema.json	JSON Schema expected by the app/local server
model-comparison-data.json	Compact benchmark summary

Base Model

• Base model: HuggingFaceTB/SmolLM2-135M-Instruct
• Base model license: Apache-2.0
• Fine-tuned model license: Apache-2.0
• Format: GGUF Q4_K_M
• Approximate GGUF size: 101 MiB

Intended Use

Use this model as an OpenAI-compatible local endpoint for Romax Pass AI:

llama-server \
  -m card-designer-q4_k_m.gguf \
  --alias card-designer-local \
  --host 127.0.0.1 \
  --port 8080 \
  -c 512 \
  -t 4 \
  -np 1

Then configure the app:

AI_PROVIDER=local
LOCAL_LLM_BASE_URL=http://127.0.0.1:8080/v1
LOCAL_LLM_MODEL=card-designer-local
LOCAL_LLM_CONCEPT_COUNT=4
LOCAL_LLM_MAX_TOKENS=430

Output Contract

The model should return one compact JSON object:

{
  "id": "local-premium-2",
  "name": "Premium Wallet Pass",
  "description": "A brand-immersive card concept.",
  "mood": "premium",
  "requiredFields": ["name", "memberId", "tier", "qrCode"],
  "designTokens": {
    "primaryColor": "#7c3f24",
    "secondaryColor": "#f0c987",
    "textColor": "#ffffff",
    "orientation": "portrait",
    "backgroundMode": "gradient",
    "colorUsage": "full",
    "brandAssetSource": "logo",
    "brandAssetTreatment": "background-emblem",
    "brandAssetIntensity": "medium",
    "usesLogo": true,
    "usesQr": true,
    "usesPhoto": false,
    "usesDecorativeArt": false
  }
}

The app validates output against local-concept.schema.json and repairs hard brand constraints before rendering.

Training

• Method: LoRA supervised fine-tuning
• Base model: HuggingFaceTB/SmolLM2-135M-Instruct
• Training data: Romax Card Design V6 synthetic dataset
• Training examples: 5,000
• Validation examples: 500
• Held-out final test examples: 500
• Training code: see the GitHub repository linked from this model card

Evaluation

Measured local V6 Q4 benchmark highlights:

Metric	Value
Request success rate	100%
Average evaluator score	73.8%
Brand color preservation	100%
Locked-field preservation	100%
Exact optional-field match	25%
Categorical design accuracy	57.3%

In a product-path benchmark, GPT-5.4 mini generated eight concepts in 8.86s. The local model generated four concepts sequentially under a strict ARM64 Docker constraint and the deterministic app filled four fallback concepts; that path completed in 130.69s.

Limitations

• Specialized for Romax Pass AI concept JSON only.
• Not a general chatbot.
• Can over-select optional fields.
• Less nuanced than larger cloud models for open-ended creative prompts.
• Needs schema-constrained decoding plus app-side validation.
• Inherits small-model limitations from SmolLM2, including possible inconsistency and bias.

GitHub

Source code, training data, training scripts, and documentation are available in the project repository.