sugiv/cardvaultplus-500m

CardVault+ SmolVLM-500M Fine-tuned Model

Model Description

CardVault+ is a specialized fine-tuned version of HuggingFaceTB/SmolVLM-500M-Instruct, optimized for credit card information extraction. This model can accurately extract structured data from credit card images in JSON format.

Key Features

• Vision-Language Processing: Processes credit card images and extracts text
• Structured Output: Returns data in consistent JSON format
• High Accuracy: Fine-tuned on 9,612 synthetic credit card images
• Production Ready: Optimized for real-world deployment

Training Details

• Base Model: HuggingFaceTB/SmolVLM-500M-Instruct
• Training Method: LoRA (Low-Rank Adaptation) fine-tuning
• Training Data: 9,612 synthetic credit card images
• Training Epochs: 4
• Final Validation Loss: 0.000056
• Training Date: July 22, 2025

Usage

from transformers import AutoProcessor, Idefics3ForConditionalGeneration
from PIL import Image
import torch

# Load model and processor
model = Idefics3ForConditionalGeneration.from_pretrained(
    "sugiv/cardvaultplus-500m",
    torch_dtype=torch.float16,
    trust_remote_code=True
)
processor = AutoProcessor.from_pretrained(
    "sugiv/cardvaultplus-500m", 
    trust_remote_code=True
)

# Process credit card image
image = Image.open("credit_card.jpg")
prompt = '''<|im_start|>user
Extract the card information from this image in the following JSON format:
{
    "card_number": "XXXX XXXX XXXX XXXX",
    "cardholder_name": "FULL NAME", 
    "expiry_date": "MM/YY",
    "cvv": "XXX",
    "card_type": "VISA/MASTERCARD/AMEX/DISCOVER"
}

<image>
<|im_end|>
<|im_start|>assistant
'''

inputs = processor(images=image, text=prompt, return_tensors="pt")
with torch.no_grad():
    generated_ids = model.generate(**inputs, max_new_tokens=200)
output = processor.decode(generated_ids[0], skip_special_tokens=True)

Expected Output Format

{
    "card_number": "4532 1234 5678 9012",
    "cardholder_name": "JOHN DOE",
    "expiry_date": "12/27", 
    "cvv": "123",
    "card_type": "VISA"
}

Technical Specifications

• Model Size: ~1GB (merged weights)
• Architecture: Idefics3ForConditionalGeneration
• Input Resolution: Variable (optimized for card images)
• Output Format: Structured JSON
• Inference Speed: ~2-3 seconds per image (RTX A6000)

Limitations and Ethical Considerations

• Intended Use: Educational and development purposes only
• Data Privacy: Do not use with real credit card data
• Security: Always implement proper data handling and security measures
• Compliance: Ensure compliance with financial data regulations

License

Apache 2.0