# **Nemotron-o1: LLM-Agent Dataflow Analysis**

### **Overview**
Nemotron-o1 is an agentic framework leveraging large language models (LLMs) for automated data flow analysis. This tool focuses on source/sink extraction, data flow summaries, and path feasibility analysis. Nemotron-o1 was benchmarked against CWE-369 test cases, demonstrating competitive performance compared to GPT-3.5 and GPT-4 in C/C++ programs.

### **Key Features**
- **Source and Sink Extraction:** Identifies critical points in the data flow for vulnerabilities.
- **Data Flow Summary:** Summarizes paths and facts for better visualization and reasoning.
- **Path Feasibility Analysis:** Uses Z3 solver for feasibility checks and detailed debugging.
- **Benchmarking:** Tested on Juliet Test Suite 1.1 with CWE-369-DBZ (Divide-by-Zero) scenarios.

### **Repository Structure**
```
Nemotron-o1-LLM-Agent-Dataflow-Analysis/
├── Phase_1_Script.py   # Source and Sink Extraction
├── Phase_2_Script.py   # Data Flow Summary Generation
├── Phase_3_Script.py   # Path Feasibility Analysis
├── .DS_Store           # Metadata file (optional, remove if not needed)
└── README.md           # Project documentation
```

### **Installation**
1. Clone the repository:
   ```bash
   git clone https://github.com/yuvrajpant56/Nemotron-o1-LLM-Agent-Dataflow-Analysis.git
   cd Nemotron-o1-LLM-Agent-Dataflow-Analysis
   ```
2. Install the required dependencies (Python 3.8+ recommended):
   ```bash
   pip install -r requirements.txt
   ```
   *(Add a `requirements.txt` file listing dependencies such as `z3-solver` and any LLM-related packages.)*

3. Set up necessary API keys or environment variables for LLMs if applicable.

### **Usage**
#### Phase 1: Source and Sink Extraction
Run the script for source/sink identification:
```bash
python Phase_1_Script.py --input your_code.c
```

#### Phase 2: Data Flow Summary
Generate a summary of the data flow paths:
```bash
python Phase_2_Script.py --input extracted_sources_and_sinks.json
```

#### Phase 3: Path Feasibility Analysis
Analyze the feasibility of paths using the Z3 solver:
```bash
python Phase_3_Script.py --input data_flow_summary.json
```

### **Benchmarking Results**
Nemotron-o1 was tested on the Juliet Test Suite 1.1 with CWE-369 benchmarks. Key results include:

| **Phase** | **Model**         | **Precision** | **Recall** | **F1 Score** |
|-----------|-------------------|---------------|------------|--------------|
| Phase 1   | Nemotron 70B      | 100%          | 100%       | 1.00         |
| Phase 2   | Nemotron 70B      | 87.73%        | 87.9%      | 0.88         |
| Phase 3   | Nemotron 70B      | 86.67%        | 90.75%     | 0.92         |

### **Future Work**
- Extend support to languages beyond C/C++.
- Conduct ablation studies to improve model performance.
- Optimize the agentic framework for faster inference times.
- Explore additional bugs beyond CWE-369.