ModelGuard 🛡️

LLM Safety, Red-Teaming & Capability Evaluation Framework

Evaluate, benchmark, and compare Large Language Models across safety and capability dimensions using automated red-team pipelines and interactive analytics dashboards.

---

🚀 Overview

ModelGuard is a comprehensive evaluation framework designed to benchmark Large Language Models (LLMs) on:

• 🔐 Safety robustness
• 🎯 Capability performance
• ⚠️ Jailbreak resistance
• 📊 Hallucination analysis
• 🤖 LLM-as-a-Judge scoring
• 📈 Interactive evaluation dashboards

The framework combines:

• capability benchmarks
• adversarial attack datasets
• automated judging pipelines
• metric aggregation
• Streamlit-based visualization

to provide end-to-end LLM evaluation workflows.

---

🧠 Evaluation Pipeline

Capability Benchmarks + Red-Team Attack Sets
                    ↓
               Target LLM(s)
                    ↓
           LLM-as-a-Judge Layer
                    ↓
      Safety Metrics | Capability Metrics
                    ↓
     Interactive Analytics Dashboard

---

✨ Features

• Multi-model evaluation support
• Automated red-team attack testing
• LLM-as-a-judge evaluation layer
• Capability benchmarking pipelines
• Safety metric computation
• Hallucination tracking
• Interactive Streamlit dashboard
• Model comparison analytics
• Config-driven architecture
• Mock evaluation support for testing

---

📊 Supported Benchmarks

Category	Benchmarks
Capability	TruthfulQA, GSM8K, MMLU, HellaSwag
Safety	HarmBench, AdvBench
Jailbreaks	DAN, Dev Mode, Base64 prompts

---

🏗️ Architecture

Four-Layer Evaluation System

Layer 1 — Inputs

• Capability benchmark datasets
• Adversarial attack datasets
• Jailbreak prompts/templates

Layer 2 — Target LLMs

Supports configurable evaluation across:

• GPT-4o
• Claude Sonnet
• Mistral-7B
• Llama-3

Layer 3 — LLM-as-a-Judge

Automated scoring for:

• Harmfulness
• Truthfulness
• Refusal quality

Layer 4 — Metrics & Analytics

Safety Metrics	Capability Metrics
Attack Success Rate (ASR)	Accuracy
Refusal Rate	BLEU
False Positive Rate	BERTScore
Per-category ASR	Hallucination Rate

---

🖥️ Dashboard Modules

Module	Description
Observe	Live evaluation monitoring
Run Evaluation	Launch evaluation runs
Safety Analysis	ASR and jailbreak analytics
Capability Analysis	Benchmark performance tracking
Model Comparison	Multi-model score comparison

---

📁 Project Structure

ModelGuard/
├── config/
│   ├── models.yaml
│   ├── benchmarks.yaml
│   └── attack_sets.yaml
│
├── data/
│   ├── capability_benchmarks/
│   ├── attack_sets/
│   └── results/
│
├── safeeval/
│   ├── models/
│   ├── benchmarks/
│   ├── attacks/
│   ├── judge/
│   ├── metrics/
│   └── pipeline.py
│
├── dashboard/
│   ├── app.py
│   ├── charts/
│   └── data_loader.py
│
├── scripts/
│   └── run_eval.py
│
└── tests/
    └── test_pipeline.py

---

⚡ Quick Start

1. Clone Repository

git clone https://github.com/maynkxx/ModelGuard.git
cd ModelGuard

---

2. Create Virtual Environment

python3 -m venv venv
source venv/bin/activate

---

3. Install Dependencies

pip install -r requirements.txt

---

4. Configure API Keys (Optional)

export OPENAI_API_KEY=your_key
export ANTHROPIC_API_KEY=your_key

Demo mode works without API keys.

---

5. Launch Streamlit Dashboard

streamlit run dashboard/app.py

---

6. Run CLI Evaluation

python scripts/run_eval.py \
    --models gpt-4o \
    --models claude-sonnet \
    --num-bench 20 \
    --num-attack 20

---

🧪 Running Tests

pytest tests/ -v

All tests use mock responses and mock judge scores.

---

📌 Future Improvements

• Additional red-team attack datasets
• Latency benchmarking support
• PDF evaluation report export
• Distributed benchmark execution
• Cost estimation dashboard
• Real-time evaluation monitoring

---

🤝 Contributing

Contributions, feature requests, and improvements are welcome.

Feel free to fork the repository and submit pull requests.

---

📄 License

This project is intended for research and educational purposes.

---

👨‍💻 Author

Mayank Choudhary

AI/ML • LLM Systems • RAG • AI Safety • Evaluation Frameworks