LEAN AND MEAN: HOW WE FINE-TUNED A MODEL FOR SECRET DETECTION IN CODE

Summary from the TLDR InfoSec newsletter

Wiz developed a small language model based on Llama 3.2 1B to detect secrets in code. It utilized Claude Sonnet 3.7 to create a training dataset of secrets based on scans of public GitHub repositories. It then fine-tuned Llama 3.2 1B using Low-Rank Adaptation (LoRA) and quantization, achieving an 82% recall rate and an 85% precision rate on commercial CPUs.