Modelpublic

SehanKim/qwen2.5-coder-security-v4-gguf

sourceHugging Faceupdated 4mo agoView on Hugging Face

1likes21downloads

Model Card

ScanOps QLoRA v4 — Security Vulnerability Detection

QLoRA fine-tuned Qwen2.5-Coder-1.5B-Instruct for CVE/CWE vulnerability analysis.

What's New in v4

—Full retrain from scratch — no catastrophic forgetting
—1,000 training samples (367 original + 633 new)
—CWE Top-25 full coverage (2023)
—CVSS base score output — new field in response
—35 CWE types across Python, Node.js, Java, React, Go, Ruby, PHP, C, GitHub Actions

Model Details

—Base: Qwen/Qwen2.5-Coder-1.5B-Instruct
—Fine-tuning: QLoRA (r=32, alpha=64, 3 epochs, cosine schedule)
—Training data: 1,000 samples, 35 CWE types, 9 languages
—Quantization: Q4KM (GGUF)
—Target detection rate: 98%+

Response Format

VULNERABILITY: CWE-89 SQL Injection
SEVERITY: CRITICAL
CVSS: 9.8
ATTACK: 공격자가 username 파라미터에 ' OR 1=1-- 를 주입해 인증을 우회합니다.
FIX:
cursor.execute("SELECT * FROM users WHERE id=%s", (user_id,))

Usage with Ollama

bash

ollama pull hf.co/SehanKim/qwen2.5-coder-security-v4-gguf:Q4_K_M