JeloH/xGenq-qwen2.5-coder-1.5b-instruct-OKI

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:1.5BQuant:BF16Ctx Length:32kPublished:Jun 26, 2025Architecture:Transformer Warm

JeloH/xGenq-qwen2.5-coder-1.5b-instruct-OKI is a 1.5 billion parameter, instruction-tuned large language model built on Qwen2.5-Coder-1.5B-Instruct, specifically designed for software security and malware analysis. It has been domain-adaptively pretrained on extensive malware datasets, including both source and assembly code, and features a 32768 token context length. This model excels at analyzing suspicious code, explaining malware behaviors, and generating structured forensic reports, making it highly effective for cybersecurity workflows.

Loading preview...

JeloH/xGenq-qwen2.5-coder-1.5b-instruct-OKI: Domain-Adapted LLM for Software Security

XGen-Q is a specialized large language model (LLM) developed by JeloH, focusing on software security and malware analysis. Built upon the Qwen2.5-Coder-1.5B-Instruct architecture, this 1.5 billion parameter model has been extensively pretrained on large-scale malware datasets, encompassing both source code and assembly code, and supports a 32768 token context length.

Key Capabilities

  • Malware Analysis: Designed to analyze suspicious code snippets.
  • Behavioral Explanation: Explains the behaviors of malware samples.
  • Forensic Reporting: Produces structured forensic reports to aid security analysts in understanding malicious intent.
  • Two-Stage Reasoning: Employs a unique two-stage reasoning pipeline that separates detailed behavioral analysis from final classification, enhancing explainability and integration into cybersecurity workflows.
  • Domain-Adaptive Pretraining: Utilizes the SBAN dataset, a multi-dimensional malware dataset, for its specialized pretraining, ensuring high relevance to software security tasks. The SBAN dataset is a key component of its training.

Performance Highlights

Evaluations demonstrate XGen-Q's superior performance in code perplexity compared to other models in its class:

  • Assembly Code Perplexity: Achieves 1.530, significantly outperforming DeepSeek-Coder-1.3B (9.183), Llama-3.1-8B-Instruct (9.972), and Phi-4-Mini (16.713).
  • Source Code Perplexity: Achieves 1.592, also showing substantial improvement over DeepSeek-Coder-1.3B (3.997), Llama-3.1-8B-Instruct (5.822), and Phi-4-Mini (7.739).

Good for

  • Cybersecurity analysts needing automated assistance in malware detection and analysis.
  • Developers integrating LLM capabilities into security tools for code auditing.
  • Research in explainable AI for software security.

Finetunes
1 models