Knowledge Graph-Enhanced LoRA Fine-Tuning for Intelligent 3GPP-Compliant 5G Network Slicing

📌 Overview

This repository demonstrates an end-to-end, highly optimized pipeline for 5G Network Slicing recommendation systems using Knowledge Graph (KG) integration and Large Language Model (LLM) fine-tuning.

Specifically, this project utilizes Parameter-Efficient Fine-Tuning (PEFT) via LoRA (Low-Rank Adaptation) on the Llama-3.2-1B-Instruct model. By converting raw telecommunication tabular data (throughput, latency, reliability) into structured Knowledge Graph triples, we train an AI agent to act as a 5G network slicing expert that maps real-world parameters to 3GPP standards (e.g., TS 23.501) and recommends optimal network configurations.

Designed by Senthilkumar Vijayakumar (IEEE Senior Member).

🚀 Key Features

• Knowledge Graph Construction: Converts raw 5G dataset parameters into Semantic Subject-Predicate-Object (SVO) triples using pandas and visualizes the ontology with NetworkX.
• LLM Prompt Engineering: Transforms graph data into structured instruction datasets optimized for Llama-3's chat template.
• LoRA & PEFT Fine-Tuning: Efficiently fine-tunes the meta-llama/Llama-3.2-1B-Instruct model using trl (SFTTrainer) and peft, drastically reducing GPU VRAM requirements.
• Apple Silicon (MPS) & CUDA Support: Dynamically handles precision and quantization. Includes stability workarounds specifically for macOS/MPS hardware.
• Inference Pipeline: Merges the base model with the trained LoRA adapter for fast, accurate inference on unseen 5G network conditions.

🧠 Technical Architecture

1. Data Processing: Parses network parameters (Throughput, Latency, Density, Mobility, Error) and tags them to specific 3GPP standards.
2. Graph Visualization: Uses NetworkX spring layouts to visualize the relationships between standards, conditions, and network slicing plans.
3. Model Initialization: Loads the Llama-3 model in FP16 (or 4-bit NF4 via bitsandbytes on CUDA devices).
4. Supervised Fine-Tuning (SFT): Employs targeted LoRA configuration on Attention (q_proj, v_proj, etc.) and MLP layers to optimize domain-specific adaptation.
5. Evaluation: Extracts training metrics and plots loss curves. Evaluates substring match accuracy against expected network recommendations.

🛠️ Installation & Setup

Ensure you have Python 3.10+ installed. Install the necessary dependencies:

pip install pandas networkx matplotlib torch transformers datasets peft trl bitsandbytes accelerate scikit-learn

Hugging Face Authentication

The Llama-3 model requires authentication. Export your Hugging Face token in your terminal before running the notebooks or scripts:

export HF_TOKEN="your_hugging_face_token_here"

📁 Repository Structure

• Comprehensive_5G_KG_LoRA.ipynb: The primary, fully-commented Jupyter Notebook detailing the entire pipeline from KG creation to model inference.
• kg_lora_llama_sft.py: Standalone Python script for automated batch processing and training.
• network_slicing_300.csv / kg_instruction_data_example.csv: Sample 5G telemetry and formatted instruction datasets.

📈 Search Engine Optimization (SEO) Keywords

5G Network Slicing, Knowledge Graph, LLM Fine-Tuning, LoRA, Llama-3, Parameter-Efficient Fine-Tuning (PEFT), Telecom AI, PyTorch, Hugging Face Transformers, AI Networking, 3GPP TS 23.501.

🤝 Contributing

Contributions, issues, and feature requests are welcome! Feel free to check the issues page.

📝 Citation

If you utilize this framework or code in your research, please use the following citation:

@software{Vijayakumar_KG_LoRA_5G_2026,
  author = {Vijayakumar, Senthilkumar},
  title = {Knowledge Graph-Enhanced LoRA Fine-Tuning for Intelligent 3GPP-Compliant 5G Network Slicing},
  year = {2026},
  url = {https://github.com/senthilv83/LLM-FineTuning},
  orcid = {0009-0009-6436-9003}
}

(See CITATION.cff for more details).