Associated Paper

This dataset release accompanies the paper Bridging the sim-to-real gap in RF localization with large-scale synthetic pretraining, published in Information Fusion. The Rome and Oslo real-world data, together with the synthetic RF localization datasets packaged here, are used in the study to train, pretrain, fine-tune, and evaluate models for sim-to-real RF fingerprinting localization.

RF-Loc-Sim2Real Dataset Export

This dataset contains RF-Loc-Sim2Real crop data packaged as compressed crop shards. It preserves the original project data layout after extraction.

Datasets included:

• oslo_a
• rome_a
• rome_bp
• rome_c

The crop files are stored in crops_shards/*.tar.zst to avoid uploading millions of small files directly to Hugging Face. Split JSON files and tabular source files are kept as normal files.

Restore

Install the lightweight restore dependencies:

python -m pip install -r scripts/requirements.txt

Download or clone this Hugging Face dataset repository, then unpack:

python scripts/unpack_dataset.py \
  --input . \
  --output /path/to/restored_rf_loc_data

Verify the restored project-compatible layout:

python scripts/verify_rf_loc_format.py \
  --root /path/to/restored_rf_loc_data

For RF-Loc-Sim2Real training configs, set dataset main paths to /path/to/restored_rf_loc_data/<dataset>/crops and split paths to /path/to/restored_rf_loc_data/<dataset>/splits.