Overview
NeuroengineAI/ZeroShot-Qwen3-14B-preview is a 14 billion parameter instruction-tuned Large Language Model developed by Alfredo Ortega, specifically fine-tuned from Qwen/Qwen3-14B. Its primary focus is to enhance automated bug hunting and code auditing capabilities, bridging the gap between smaller, faster models and the high-reasoning demands of vulnerability research.
Key Capabilities & Specialization
- Enhanced Bug Hunting: Fine-tuned with over 10,000 thinking traces of public CVEs (Common Vulnerabilities and Exposures), totaling nearly 500MB of text.
- Performance Improvement: Benchmarks indicate approximately a 20% enhancement in bug hunting capabilities over the base Qwen3-14B model.
- Efficiency for Code Auditing: Designed to be efficient and cost-effective for analyzing large, enterprise-scale codebases, offering a specialized solution without the high costs or slow processing of larger, general-purpose models.
- Multilingual Support: The model supports multiple languages for NLP tasks.
Benchmarks
On the CrashBench benchmark, the Zeroshot-Qwen3-14B-preview achieved a score of 68.8, significantly outperforming the base Qwen3-14B's score of 55.6. This demonstrates the effectiveness of its specialized fine-tuning even with a relatively small dataset.
Recommended Use Cases
- Automated Security Analysis: Ideal for identifying and analyzing security vulnerabilities in code.
- Code Auditing: Suitable for auditing large codebases where efficiency and specialized reasoning are critical.
- Vulnerability Research: Provides high-reasoning capabilities for complex vulnerability detection tasks.