car-damage-assessment-llama-vision

Car Damage Assessment Model

This model is a fine-tuned version of meta-llama/Llama-3.2-11B-Vision-Instruct specialized for professional car damage assessment.

Model Description

This AI model can analyze car images and provide detailed damage assessments including:

• Damage Detection: Identifies specific damaged parts (bumper, hood, doors, etc.)
• Severity Classification: Categorizes damage as minor, moderate, or major
• Professional Reports: Generates insurance-quality damage descriptions
• Multi-language Support: Works with English and French terminology

Training Details

• Base Model: Llama-3.2-11B-Vision-Instruct (11B parameters)
• Training Method: LoRA (Low-Rank Adaptation) fine-tuning
• Dataset: 5,600 car damage assessment samples
• Training Time: 138 minutes on NVIDIA L40S
• Final Loss: 0.1478
• Trainable Parameters: 67,174,400 out of 10,737,395,235 total

Usage

from unsloth import FastVisionModel
import torch

# Load model and tokenizer
model, tokenizer = FastVisionModel.from_pretrained("Kakyoin03/car-damage-assessment-llama-vision")

# Prepare image and prompt
from PIL import Image
image = Image.open("car_damage.jpg")
instruction = "You are a professional car damage assessment expert. Analyze this image and provide a detailed damage report."

messages = [
    {"role": "user", "content": [
        {"type": "image"},
        {"type": "text", "text": instruction}
    ]}
]

# Generate assessment
input_text = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
inputs = tokenizer(image, input_text, return_tensors="pt")

with torch.no_grad():
    output = model.generate(**inputs, max_new_tokens=200)

assessment = tokenizer.decode(output[0], skip_special_tokens=True)
print(assessment)

Applications

• Insurance Claims: Automated damage assessment for insurance companies
• Auto Repair: Quick damage evaluation for repair shops
• Car Dealerships: Pre-purchase vehicle condition reports
• Mobile Apps: Real-time car damage assessment tools

Performance

The model demonstrates excellent performance in:

• Accurate damage part identification
• Reliable severity classification
• Detailed professional descriptions
• Multi-language damage terminology
• Fast inference (< 1 minute per assessment)

License

This model is released under the Llama 3.2 license. Please ensure compliance with the original license terms.

Acknowledgments

• Unsloth: For efficient training framework
• Meta: For the base Llama-3.2-Vision model
• Dataset: KHAOULA-KH/CAR_DOMMAGE_ for training data

Developed with

• Unsloth Framework
• PyTorch 2.4.0+cu121
• Transformers Library
• NVIDIA L40S GPU (44.7GB VRAM)

---

Trained on August 08, 2025 using advanced vision-language fine-tuning techniques.