TWC Frontier-Track OTA Dataset v1

This dataset contains over-the-air 5G NR measurements collected with the Frontier-Track prototype system. It includes the validated P3/P4 matched experiments, the P1 UE03 calibration tables, the P2 three-UE anchor evidence, and selected radio and resource traces used to check the derived tables.

The measurements come from a laboratory setup with an OAI gNB, a USRP X310, and an Open5GS 5G Core. Commercial handsets were used as UEs. The radio action range is NR MCS 14--26. Link changes refer to the OAI software TX-attenuation command; this release does not include a separate programmable-attenuator readback.

What is included

Configuration	Rows	Purpose
tb_level (p3, p4)	90,000	One validated formal new-TB record per row; P3 has 30,000 rows and P4 has 60,000 rows.
run_summary	50	One summary row per selected half-run.
pair_metrics	25	Matched Frontier/OLLA pair metrics; differences are Frontier minus OLLA.
change_summary	130	Profile-change and recovery summaries.
p1_calibration_*	1,936 points + 13 curves + checks	UE03 calibration evidence and monotone MCS crossing fits. Each table is a separate config because the schemas differ.
p2_*	50 evidence rows + 18 crossings + checks	UE02/UE03/UE04 anchor evidence and boundary-limited crossing results. Each table is a separate config because the schemas differ.

The relationship indexes (block_manifest, run_index, valid_pairs) and the three-UE label table (terminals) are available as separate configs as well.

The artifacts/ directory contains selected raw radio, metric, resource and acceptance files for the 50 P3/P4 half-runs, plus the P1/P2 supporting traces. The source/ directory contains selected sanitized CSV evidence and release metadata. Parquet is the recommended format for analysis. CSV copies are included for auditing and for readers that do not use the Hugging Face tooling.

Important interpretation limits

• These are engineering measurements from one OAI/USRP laboratory setup. They are not a commercial-base-station benchmark or a channel-sounding dataset.
• P1 MCS25/26 include boundary-truncated fits. P2 does not identify a single exact terminal correction shared by all three UEs.
• P3 uses the 20→16 step profile. P4 uses the 22→16→22→16 profile.
• Use the matched pair as the statistical unit when comparing Frontier and OLLA. Treating the two controllers as unrelated groups changes the analysis.
• The results are metric-specific. The dataset does not imply that Frontier is better on every metric; use data/pair_metrics and data/metric_summary to reproduce the reported comparisons.

Privacy and safety

The public tree excludes IMSI/SUPI, IMEI, USB serial numbers, ADB mappings, Open5GS journals, private gNB/core configurations, host launchers, and local absolute paths. UE aliases and handset model codes are used only as experiment labels. No credentials or private keys are included.

metadata/release_manifest.json records the exclusion policy. SHA256SUMS covers all payload, documentation and repository-control files; it necessarily excludes itself and the generated public-file manifest.

Loading the data

After uploading this directory to a Hugging Face dataset repository, the main table can be loaded with:

from datasets import load_dataset

repo_id = "xiaoxiao720/Frontier-Track"
tb = load_dataset(
    repo_id,
    data_files={
        "p3": "data/tb_level/p3-*.parquet",
        "p4": "data/tb_level/p4-*.parquet",
    },
)
print(tb["p3"].num_rows, tb["p4"].num_rows)

For a quick paired analysis, load data/pair_metrics/p3-*.parquet and data/pair_metrics/p4-*.parquet. The complete list of public files and their hashes is in metadata/public_file_manifest.csv.

Tables with different column schemas are intentionally separate Hugging Face configs. For example:

from datasets import load_dataset

p1_points = load_dataset(
    "xiaoxiao720/Frontier-Track",
    name="p1_calibration_points",
)
p1_curves = load_dataset(
    "xiaoxiao720/Frontier-Track",
    name="p1_calibration_curves",
)

Citation

See CITATION.cff.