Skip to main content
Search

# Sigma Rules

1 related articles

Subscribe to updates
  1. arXiv

    Sigma-Hunter fine-tunes a 7B model on 3,635 Sigma rules, scoring 8.17 overall against 7.88 for the strongest general-purpose baseline

    The work builds Sigma-Hunter by expanding 3,635 validated open-source Sigma rules into 7,663 question-answer and analyst-reasoning instruction examples, partitioned at the source-rule level to prevent leakage; after LoRA fine-tuning of a 7B Mistral and a Phi-4 model, evaluation on 188 held-out rule-generation examples across syntax, approximate field consistency, and semantic judging gives Sigma-Hunter-Mistral an overall 8.17, above the strongest general-purpose baseline at 7.88 and untuned Mistral at 4.61.

You’re all caught up