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

1. Introduction

XGen-Q is a domain-adapted large language model designed for software security analysis and malware analysis.

The model is built on Qwen2.5-Coder-1.5B-Instruct and further pretrained on large-scale malware datasets containing both source code and assembly code. XGen-Q is designed to analyze suspicious code, explain malware behaviors, and produce structured forensic reports that help security analysts understand why a sample may be malicious.

A key feature of XGen-Q is its two-stage reasoning pipeline, which separates detailed behavioral analysis from the final classification decision. This design improves explainability and makes the model suitable for integration into real-world cybersecurity workflows.

The model is domain-adaptively pretrained using the SBAN dataset, a multi-dimensional malware dataset designed for LLM pretraining in software security.

SBAN Dataset:
https://ieeexplore.ieee.org/document/11392071/

---

2. Evaluation Results

---

3. Citation

@article{jelodar2025xgenq, title={XGen-Q: An Explainable Domain-Adaptive LLM Framework with Retrieval-Augmented Generation for Software Security}, author={Jelodar, Hamed and Meymani, Mohammad and Razavi-Far, Roozbeh and Ghorbani, Ali}, journal={arXiv preprint arXiv:2510.19006}, year={2025} }

---

4. How to Use

Below is a simple example showing how to run XGen-Q using the Hugging Face Transformers library.

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

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

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

code_snippet = """
void inject_polymorphic_dll(DWORD pid) {
    // suspicious DLL injection example
}
"""

prompt = f"""
You are a cybersecurity malware analyst.

Analyze the following code and produce:

1. Conclusion
2. Reasoning
3. Evidence
4. Suspicious Behavior Explanation
5. Final Judgment (malware / benign / partially malicious)

Code:
{code_snippet}
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=400,
    temperature=0.2
)

result = tokenizer.decode(outputs[0], skip_special_tokens=True)

print(result)