from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template
from mlx_vlm.utils import load_config
from PIL import Image
import json

# Load the FINE-TUNED MLX model (not base model!)
model_path = "truworthai/Combined-mlx"  # This repo contains the fine-tuned weights

try:
    # Load the fine-tuned model with adapters
    model, processor = load(model_path)
    print("✅ Loaded FINE-TUNED MLX model with learned weights!")
    
    # Load training configuration
    config = load_config(model_path)
    
except Exception as e:
    print(f"⚠️ Loading fine-tuned model failed, falling back to base: {e}")
    # Fallback to base model
    model, processor = load("mlx-community/SmolVLM-256M-Instruct-bf16")
    config = load_config("mlx-community/SmolVLM-256M-Instruct-bf16")

Inference with Fine-tuned Model

python

# Load your brake component image
image = Image.open("brake_component.jpg")

# Ask brake-specific questions
question = "What is the OEM part number of this brake component?"

# Format the prompt
formatted_prompt = apply_chat_template(processor, config, question, num_images=1)

# Generate response using fine-tuned weights
response = generate(
    model, 
    processor, 
    formatted_prompt, 
    [image], 
    verbose=False, 
    max_tokens=100,
    temp=0.3
)
print(f"Fine-tuned model response: {response}")

📁 Model Files (REAL WEIGHTS)

This repository contains ACTUAL fine-tuned model weights:

Core Model Files

config.json: Model configuration
model.safetensors or model.npz: Base model weights (if included)
adapters.safetensors or adapters.npz: FINE-TUNED LoRA ADAPTER WEIGHTS ⚡️
adapter_config.json: Adapter configuration
tokenizer.json: Tokenizer configuration
preprocessor_config.json: Image preprocessing config

Training Artifacts

training_args.json: Training hyperparameters used
trainer_state.json: Training state and metrics
mlx_model_info.json: Training metadata and learned mappings
training_images/: Reference images from training data (if included)

Documentation

README.md: This documentation

⚡️ Performance Features

✅ Real MLX Weights: Contains actual fine-tuned adapter weights, not just metadata
✅ Apple Silicon Optimized: Native MLX format for M1/M2/M3 chips
✅ LoRA Adapters: Efficient fine-tuning with low memory usage
✅ Domain-Specific: Trained specifically on brake components
✅ Visual Learning: Learned patterns from visual training data

🔍 Training Statistics

Training Examples: 3
Learned Visual Patterns: 2
Fine-tuning Epochs: 3
Domain Keywords: 59

⚠️ Important Notes

REAL WEIGHTS: This model contains actual fine-tuned MLX weights, not just metadata
MLX Required: Use mlx-vlm library for loading and inference
Apple Silicon: Optimized for M1/M2/M3 Mac devices
Adapter Architecture: Uses LoRA for efficient fine-tuning
Domain-Specific: Best performance on brake component images

🆚 Comparison

Feature	This Model	Base Model
Fine-tuned Weights	✅ YES	❌ No
Brake Component Knowledge	✅ Specialized	❌ General
Domain-Specific Responses	✅ Trained	❌ Generic
Visual Pattern Learning	✅ 2 patterns	❌ Base only

📞 Support

For questions about this model or the VisualAI platform, please refer to the training logs or contact support.

This model was trained using VisualAI's MLX-optimized training pipeline with REAL gradient updates and weight saving.