FccIQ Dataset 01 - Detailed Methodology
Adaptive AI Model Partitioning Paper

This dataset is collected from a live 5G NR testbed built on the NVIDIA Aerial CUDA-accelerated RAN. It includes both Layer-1 IQ samples and Layer-2/3 radio measurements under various interference conditions such as UE-to-BS co-channel interference, BS-to-BS TDD mismatch, and external jamming.

Uplink throughput was benchmarked using iPerf3 across high-load and low-load configurations.

Data Collection Methodology:

  • Layer 1 I/Q Samples: Collected and pushed to ClickHouse database autonomously by cuBB
  • Layer 2/3 Metrics: Collected using xApp monitoring script xapp_mac_rlc_pdcp_gtp_moni.py
  • Architecture: xApp-based monitoring system using RIC (RAN Intelligent Controller)
  • Data Storage: ClickHouse database for high-performance time-series data storage
  • Collection Rate: 20 Hz (50ms intervals) for real-time monitoring

Layer 1 I/Q Samples:

  • Collection Method: Automatically collected and pushed to ClickHouse by cuBB (NVIDIA CUDA-accelerated baseband)
  • Baseband I/Q samples from the 5G NR air interface
  • Time-frequency domain representations

Collected Metrics by Protocol Layer:

  • Physical Layer: RSRP, SS-RSRP, SS-RSRQ, SS-SINR, CQI
  • MAC Layer: BLER, HARQ retransmission statistics, MCS, PHR
  • RLC Layer: Buffer status, PDU transmission metrics, retransmission statistics
  • PDCP Layer: Buffer status, throughput measurements, sequence number tracking
  • GTP Layer: Tunnel statistics, packet transmission rates, establishment events

Interference Scenarios:

  • UE-to-BS Co-channel Interference: Multiple UEs operating on the same frequency
  • BS-to-BS TDD Mismatch: Base station timing misalignment scenarios
  • External Jamming: Intentional interference injection for robustness testing

The dataset is designed to support engineering and research tasks in:

  • ML-based throughput estimation using combined I/Q and KPI features
  • Adaptive model partitioning between UE and edge servers
  • Performance analysis of 5G NR under controlled noise and interference scenarios
  • Privacy-aware edge intelligence and resource optimization

The data is stored in ClickHouse with the following table structure:

  • Layer 1 I/Q Samples: fapi , fh
  • Layer 2/3 Metrics: MAC_KPIs_2, RLC_KPIs, PDCP_KPIs, GTP_KPIs
This provides a realistic foundation for building and benchmarking AI/ML pipelines that operate in dynamic, interferenced 5G environments.

📄 Research Paper

Read our published research on adaptive AI model partitioning

PDF
Tam Thanh Nguyen, Tuan Van Ngo, Long Thanh Le, Yong Hao Pua, Mao Van Ngo, Binbin Chen, and Tony Q. S. Quek, "Adaptive AI Model Partitioning over 5G Networks," In Proceeding of the IEEE Global Communications Conference (GLOBECOM), Taipei, Taiwan, December 2025 (to appear);

💻 Source Code

Access the xApp monitoring script and tools

GitHub Repository

📥 Download Dataset

Access the complete dataset with all collected metrics

Download