Local Llama 1B beats Wiz secret scan

The real takeaway is not just “small model good,” but that careful local data work can turn a 1B model into a surprisingly strong security classifier without cloud APIs.

• –Structured JSON output was a key unlock; the author had to train past early schema-compliance failures before the model became usable in production-like workflows
• –Label quality mattered as much as architecture: removing a bad high-entropy class and fixing mislabeled “negative” samples that were actually passwords improved recall
• –Using Qwen3-Coder-Next for local labeling is the most interesting part of the pipeline, because it shows one specialized local model bootstrapping another
• –Bigger Qwen 3.5-2B and 4B variants outperformed the 1B baseline, but the cost moved to VRAM pressure and slower inference rather than better efficiency
• –This is a strong proof point for privacy-sensitive security tooling: local fine-tunes can be good enough to complement or even outperform legacy regex-heavy detectors