Leopo1d/OpenVul-Qwen3-4B-SFT-ep5

TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Feb 14, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Leopo1d/OpenVul-Qwen3-4B-SFT-ep5 is a 4 billion parameter Qwen3-based language model developed by Youpeng Li, Fuxun Yu, and Xinda Wang, fine-tuned for vulnerability detection in C/C++ code. It specializes in context-level security analysis, utilizing inter-procedural information rather than isolated functions. The model is trained on high-quality vulnerability reasoning Chain-of-Thought (CoT) data using rejection sampling to prevent ground-truth leakage and reasoning hallucinations. Its primary use is identifying security flaws and Common Weakness Enumeration (CWE) standards in code.

Loading preview...

OpenVul-Qwen3-4B-SFT-ep5: Vulnerability Detection Model

This model, developed by Youpeng Li, Fuxun Yu, and Xinda Wang, is a 4 billion parameter Qwen3-based language model specifically fine-tuned for vulnerability detection in C/C++ code. It leverages Chain-of-Thought (CoT) data for security reasoning and instruction-following capabilities.

Key Capabilities

  • Context-level Vulnerability Detection: Unlike models that analyze isolated functions, OpenVul-Qwen3-4B-SFT-ep5 focuses on inter-procedural contexts, incorporating global variables, type definitions, and callee functions for a more comprehensive analysis.
  • High-Quality Training: The model was trained using Rejection Sampling distilled from DeepSeek-R1-0528, a method chosen to prevent "ground-truth leakage" and reasoning hallucinations during the fine-tuning process.
  • CWE Standard Focus: It is designed to identify security flaws and classify them according to Common Weakness Enumeration (CWE) standards.

Good For

  • Developers and security researchers needing to identify security vulnerabilities in C/C++ codebases.
  • Automated code review systems focused on security analysis.
  • Environments requiring precise, evidence-based vulnerability analysis without speculation.

For optimal inference, the developers recommend using vLLM with specific parameters such as enable_thinking=True, n=8, repetition_penalty=1.0, temperature=0.6, top_p=0.95, top_k=20, and max_tokens=32768.