URL extraction results: testing on-premise LLMs on selected exploits
Description
This dataset presents URL extraction results from testing three on-premise Large Language Models (LLMs): Llama 3.1: 8B, Qwen 2.5 Coder: 7B Instruct, and Codestral: 22B. Each model has its directory containing subdirectories organized by CVE names, which in turn contain text files named after corresponding exploit IDs from ExploitDB or PacketStorm. These text files represent each model's output when tasked with extracting URLs that would be accessed during the execution of the exploit code, enabling direct reference to the original exploit samples.
Files
Steps to reproduce
This dataset was generated in September 2024, using the Ollama Python library on a system with 32GB of VRAM. Each of the three LLMs (Llama 3.1: 8B, Qwen 2.5 Coder: 7B Instruct, and Codestral:22B) was prompted to identify URLs that would be accessed during the execution of selected exploits code, with responses saved in an organized directory structure. All models were run with default parameters, without any custom configuration or fine-tuning.
Institutions
- NASK Instytut Badawczy