"Balanced 9-Class Instruction-Formatted Subset of CIC-IoT2023 for IIoT Intrusion Detection"

"This dataset is a curated, balanced subset of the CIC-IoT2023 dataset, specifically reformatted for supervised fine-tuning of Large Language Models (LLMs) in Industrial IoT (IIoT) environments. It contains 9,000 flow-level records spanning 9 balanced categories (1,000 samples per class): Benign Traffic, DDoS (ICMP\/TCP\/UDP Flood), DoS (TCP\/UDP Flood), MITM-ARP Spoofing, and Mirai (Greeth\/UDPPlain Flood) .Unlike conventional numerical datasets, each instance is transformed into an instruction-input-output format. This structure embeds 80+ numerical flow features into natural-language prompts, enabling models like LLaMA-2 to perform domain-specific intrusion classification through instruction-driven learning. This dataset serves as the experimental foundation for the LLaMA-RAG framework, supporting research into explainable and transparent intrusion detection for smart factories."