ModelAuditor Agent

AI-powered model auditing agent with multi-agent debate for robust evaluation of machine learning models.

Setup

This repository has been tested extensively with Python 3.10.15. Typical install time via uv is less than a minute.

Using uv (recommended)

uv sync
uv run python main.py --model resnet50 --dataset CIFAR10 --weights path/to/weights.pth

Using pip

pip install -e .
python main.py --model resnet50 --dataset CIFAR10 --weights path/to/weights.pth

Medical AI dependencies (optional)

uv sync --extra medical  # or pip install -e ".[medical]"

Data and Model Downloads

Pre-trained Models (HuggingFace)

Download pre-trained ResNet50 models from HuggingFace:

# Install huggingface_hub if needed
pip install huggingface_hub

# Download all models
python -c "from huggingface_hub import hf_hub_download; [hf_hub_download('lukaskuhndkfz/ModelAuditor', f'{name}_resnet50_1_224.pt', local_dir='models') for name in ['camelyon17', 'chexpert', 'ham10000']]"

Or download individually:

# Camelyon17 (pathology)
huggingface-cli download lukaskuhndkfz/ModelAuditor camelyon17_resnet50_1_224.pt --local-dir models

# CheXpert (chest X-ray)
huggingface-cli download lukaskuhndkfz/ModelAuditor chexpert_resnet50_1_224.pt --local-dir models

# HAM10000 (dermatology)
huggingface-cli download lukaskuhndkfz/ModelAuditor ham10000_resnet50_1_224.pt --local-dir models

HAM10000 Dataset

1. Download the dataset from Harvard Dataverse:

   • HAM10000_images_part_1.zip
   • HAM10000_images_part_2.zip
   • HAM10000_metadata.tab as CSV

2. Extract and organize:

# Create data directory
mkdir -p data/ham10000

# Extract images (both parts) into data/ham10000/
unzip HAM10000_images_part_1.zip -d data/ham10000/
unzip HAM10000_images_part_2.zip -d data/ham10000/

# Copy metadata
cp HAM10000_metadata.csv data/ham10000/

3. Split into vidir_modern and rosendahl subsets:

python setup_ham10000.py

This creates the following structure:

data/ham10000/
├── vidir_modern/
│   ├── bkl/  (475 images)
│   └── mel/  (680 images)
└── rosendahl/
    ├── bkl/  (490 images)
    └── mel/  (342 images)

Usage

General Usage

python main.py --model resnet50 --dataset CIFAR10 --weights models/model.pth

Medical Models

# ISIC skin lesion classification
python main.py --model siim-isic --dataset isic --weights models/isic/model.pth

# HAM10000 dataset
python main.py --model deepderm --dataset ham10000 --weights models/ham10000.pth

Medical Example (DeepDerm)

The auditor can be tested with the DeepDerm classifier which can be downloaded here. All that is needed is a valid Anthropic API Key as can be seen below (see section 'Environment Variables') as well as a test dataset such as HAM10000.

python main.py --model deepderm --dataset ham10000 --weights models/deepderm_isic.pth

Alternatively, use our pre-trained ResNet50 model from HuggingFace (see "Data and Model Downloads" above):

python main.py --model resnet50 --dataset ham10000 --weights models/ham10000_resnet50_1_224.pt

Expected runtime varies depending on user response speed and subset time but should take less than 10 minutes in total.

Toy Example (CIFAR10)

We also prepared a small toy model, trained on CIFAR10 so the Auditor can be evaluated on natural images. All that is needed is a valid Anthropic API Key as can be seen below (see section 'Environment Variables').

python main.py --model resnet18 --dataset CIFAR10 --weights examples/cifar10/cifar10.pth

Expected runtime varies depending on user response speed and subset time but should take less than 10 minutes in total.

Options

• --subset N: Use N samples for faster evaluation
• --no-debate: Disable multi-agent debate
• --single-agent: Use single agent instead of multi-agent debate
• --device: Specify device (cpu, cuda, mps)

Environment Variables

Set your API keys:

export ANTHROPIC_API_KEY="your-key"
export OPENAI_API_KEY="your-key"  # if using non-Anthropic models

Project Structure

• main.py - Interactive model auditor with multi-agent debate
• testbench.py - Automated evaluation script
• utils/agent.py - Multi-agent conversation system
• architectures/ - Custom model architectures
• prompts/ - System prompts for different evaluation phases
• models/ - Pre-trained model weights
• results/ - Evaluation results and conversation logs