🦅 Horus-OSINT

A Cloud-Based Conversational Chatbot for Open-Source Intelligence (OSINT) and Global Threat Analysis using Fine-Tuned Lightweight LLMs.

---

📚 Model Description

• Authors:
  • Mahmoud Alyosify (Group 23 - CISC 886, Queen's University)
  • Sondos Omar
  • Mirna Embaby
• Project Context:
  Academic deliverable for CISC 886 – Cloud Computing, Queen's University.
• Base Model:
  unsloth/Meta-Llama-3-8B-Instruct-bnb-4bit
• Architecture:
  Llama-3 (8 Billion Parameters)
• Quantization/Format:
  4-bit Quantized (q4_k_m), GGUF format
• Fine-tuning Technique:
  PEFT (QLoRA) using Unsloth and HuggingFace TRL SFTTrainer
• Language:
  English
• Main Applications:
  Open-Source Intelligence (OSINT), Geopolitical Threat Analysis, Structured Military Reporting

---

🚀 Intended Uses

• Threat Intelligence Analysis:
  Query historical geopolitical events and regional instability from indexed records.
• Automated Reporting:
  Generates structured intelligence reports with clear sections: GEOPOLITICAL CONTEXT and THREAT ASSESSMENT.
• Situational Awareness:
  Analyze patterns of activities, severity levels, and trends according to historical data.

---

⚠️ Limitations & Ethical Considerations

• No Real-Time Awareness:
  All model outputs are restricted to knowledge present in training datasets (GTD & GDELT up to their latest records).
• Hallucinations:
  Outputs may contain errors or outdated information inherent to LLMs.
• Not Suited for Critical Security Decisions:
  This model is strictly for research and academic purposes – any use in security, military, or life-critical contexts requires thorough expert validation and oversight.

---

🏗️ Training & Data Pipeline Details

• Big Data Engineering:
  Sourced over 20M+ records from the Global Terrorism Database (GTD) and Global Database of Events, Language, and Tone (GDELT).
• ETL Pipeline:
  Deployed Apache Spark jobs via AWS EMR for cleaning and filtering large-scale event data.
• Data Distillation:
  Refined into an instruction-following dataset with 159,826 quality samples.
• Compute Environment:
  Fine-tuned on Google Colab T4 GPU (15–25 minutes runtime expected).
• Cloud Storage:
  Model artifacts are versioned on Amazon S3:
  s3://horus-25bbdf-g23-bucket/models/horus-llama3-osint-Q4_K_M.gguf

---

💻 Usage Example

Horus-OSINT is exported as a compact GGUF model, ready out-of-the-box for Ollama:

Running Locally with Ollama

1. Download the llama-3-8b-instruct.Q4_K_M.gguf file.
2. Place a file named Modelfile in the same directory:

   FROM ./llama-3-8b-instruct.Q4_K_M.gguf
   

3. In your terminal, run:

   ollama create horus-osint -f Modelfile
   ollama run horus-osint
   

---

🏷️ Citation

If you use this model in your research or applications, please cite:

@misc{horus_osint_2026,
  author = {Mahmoud Alyosify and Sondos Omar and Mirna Embaby},
  title = {Horus-OSINT: Cloud-Based OSINT and Threat Analysis using LLMs},
  year = {2026},
  publisher = {Hugging Face},
  institution = {Queen's University},
  url = {https://huggingface.co/mahmoudalyosify/Horus-OSINT}
}

---

📜 License

Apache 2.0 (subject to Meta’s Llama 3 Acceptable Use Policy)

---

📬 Contact

For questions or feedback, connect with the development team:

• Mahmoud Alyosify
• Sondos Omar
• Mirna Embaby

---

All technical details and contributor links are included for thorough documentation and professional presentation. Simply copy and use as your model card on Hugging Face!