hayulalab/Circuit-Aware-SAIF-7B

Circuit-Aware SAIF-7B

سَيْف — "sword" in Arabic

LoRA adapter for Qwen2.5-7B, trained with Circuit-Aware methodology using J-lens (Global Workspace Theory) to reduce hallucinations in security vulnerability analysis.

Training Details

• Base Model: Qwen/Qwen2.5-7B
• Rank: 32
• Iterations: 1000
• Batch Size: 4
• Learning Rate: 1e-5
• Val Loss: 2.051
• Train Loss: 0.043

Dataset

430 Arabic-English security samples across 9 domains: WebApp, Network, Reverse Engineering, Exploitation, Kernel, Browser, Vulnerability Discovery, Bug Bounty, Exploiter

Circuit-Aware Methodology

This adapter was trained using J-lens circuit mapping, which analyzes model activation patterns across 28 transformer layers to identify and strengthen factual reasoning circuits while suppressing hallucination-prone pathways.