QUELL: An On-Device Benchmark of Quantized Edge LLMs versus Classical Machine Learning for IoT Intrusion Detection
QUELL — reproducibility package (v4.0). This repository accompanies the paper “QUELL: An On-Device Benchmark of Quantized Edge LLMs versus Classical Machine Learning for IoT Intrusion Detection.” QUELL fine-tunes decoder-only LLMs (GPT-2-medium, Qwen2.5-0.5B/1.5B, Llama-3.2-1B) on network-flow records serialized as text, quantizes the Qwen2.5-1.5B detector, deploys it on an NVIDIA Jetson Orin Nano Super (GGUF/llama.cpp), and compares it with random-forest and XGBoost detectors on Edge-IIoTset, CICIoT2023 and N-BaIoT under a leakage-controlled, multi-seed protocol. Axes: detection quality, on-device latency/energy/memory, quantization retentio
Resource information
- Format
- zip
- Original repository
- 23448283
- DOI
- 10.5281/zenodo.23448283
- Credit basis
- Creators in Zenodo record metadata
- Access
- Open files available from the original Zenodo deposit
- Last checked
- Nov 5, 2026
- License
- mit-license
- Last record update
- Nov 5, 2026
About this resource
QUELL — reproducibility package (v4.0). This repository accompanies the paper “QUELL: An On-Device Benchmark of Quantized Edge LLMs versus Classical Machine Learning for IoT Intrusion Detection.” QUELL fine-tunes decoder-only LLMs (GPT-2-medium, Qwen2.5-0.5B/1.5B, Llama-3.2-1B) on network-flow records serialized as text, quantizes the Qwen2.5-1.5B detector, deploys it on an NVIDIA Jetson Orin Nano Super (GGUF/llama.cpp), and compares it with random-forest and XGBoost detectors on Edge-IIoTset, CICIoT2023 and N-BaIoT under a leakage-controlled, multi-seed protocol. Axes: detection quality, on-device latency/energy/memory, quantization retention, unknown-attack generalization, random-noise sensitivity and an attacker-constrained HopSkipJump evaluation. Contents: numbered pipeline ( scripts/ , steps 00–23), raw result files of every run ( results/v4/ , with PROVENANCE.md mapping each file to its script, settings, log and manuscript table), raw run logs ( logs/v4/ ), split indices ( splits/ ), SHA-256 manifest of the raw data ( data/MANIFEST.md ), and reproduce.sh , which regenerates Tables II–XI and Figs. 4–6 from the result files and checks every reported number (103 checks). The benchmark datasets are public and not redistributed here.