QCRI/LLMxCPG-D

LLMxCPG-D

Model Description:

LLMxCPG-D is a highly effective vulnerability detection model. It is a fine-tuned version of the QwQ-32B-Preview model, optimized for a binary classification task.

This model is the second phase of the LLMxCPG framework. It takes as input a concise code slice that has been generated by the LLMxCPG-Q model and the Joern static analysis tool. The model then classifies this code slice as either 'VULNERABLE' or 'SAFE'.

How it Works:

By focusing on small, vulnerability-relevant code slices rather than entire codebases, LLMxCPG-D can make more accurate and robust predictions. This approach significantly reduces noise and allows the model to learn the fundamental characteristics of vulnerabilities, leading to superior performance on a variety of datasets.

GitHub Repository:

For more information, please visit the official GitHub repository: https://github.com/qcri/llmxcpg