models of Smart-LLaMA-DPO

This repository provides the weights for the Smart-LLaMA-DPO model, with the related research paper accepted to ISSTA 2025. Since the original training data contained content from private GitHub repositories, the model version published here has been trained using desensitized data to protect privacy and intellectual property.

We offer two versions of the model:

• base: Version containing MU (Machine Unauditable) vulnerabilities
• base_wo_mu: Version without the MU vulnerabilities

Researchers should choose base_wo_mu for experiments involving reentrancy, timestamp dependency, integer overflow/underflow, and delegatecall vulnerabilities. For research focused on Machine Unauditable vulnerabilities, the base version is recommended.