---
pretty_name: RepresentationLearning ORACLE triplets
tags:
- wireless
- rf-fingerprinting
- channel-estimation
- cfo-estimation
- iq-samples
size_categories:
- 100K<n<1M
---

# RepresentationLearning dataset

Preprocessed WiFi packets for three wireless tasks, used to train and attack the
multi-task representations in
[RepresentationLearning](https://github.com/nasimsoltani/RepresentationLearning).
Pretrained models and the run registry are in
[`Aadharsh/RepresentationLearning-models`](https://huggingface.co/Aadharsh/RepresentationLearning-models).

The packets are derived from the [ORACLE RF fingerprinting dataset](https://genesys-lab.org/oracle)
(16 USRP X310 transmitters recorded at 11 distances, 2–62 ft).

## Files

| File | Size | Contents |
|---|---|---|
| `OracleDatasetProcessed-arranged.tar.gz` | 3.3 GB | 416,244 `.mat` files (see below) |
| `rf_partition_dict_0.5.pkl` | 5.8 MB | train/val/test split and CFO normalisation constants |

## Layout

```bash
tar xzf OracleDatasetProcessed-arranged.tar.gz
```

The archive has a doubled top-level directory. The `.mat` files are in
`OracleDatasetProcessed-arranged/OracleDatasetProcessed-arranged/`; point
`DATA_BASE_PATH` there.

One sample is a **triplet** sharing the suffix `_run1_Radio<r>_<d>ft_<i>.mat`:

| Prefix | Variables | Task |
|---|---|---|
| `RFfingerprinting` | `Packet` (4000×1 complex), `Radio` (label, `"Radio0"`–`"Radio15"`) | RF fingerprinting (16 classes) |
| `CFOEstimation` | `LSTF` (160×1 complex), `CFO` (scalar, Hz) | CFO estimation |
| `ChannelEstimation` | `LLTF` (160×1 complex), `EstChannnel` (52×1 complex) | Channel estimation |

`EstChannnel` really is spelled with three n's. The archive also contains 24,525
`CFOEstimation` files without RF/channel partners; no split references them.

## Split

`rf_partition_dict_0.5.pkl` is a dict:

| Key | Value |
|---|---|
| `train` / `val` / `test` | lists of `RFfingerprinting_*.mat` filenames: 91,323 / 13,070 / 26,180 |
| `mean_cfo`, `std_cfo`, `max_cfo` | CFO statistics of the training split, used to normalise CFO targets |

Filenames are relative to `DATA_BASE_PATH`. The registry's test metrics for runs
from July 2025 onward were computed on this split, so use it unchanged when
comparing with the registry.

## Loading

```bash
git clone https://github.com/nasimsoltani/RepresentationLearning.git
cd RepresentationLearning && cp .env.example .env   # set DATA_BASE_PATH and PKL_FILE_PATH
python code/rep_lr/inference.py                     # evaluate the best released model
```

Each input is normalised to unit RMS and split into I/Q channels by
`code/rep_lr/py_datasets.py`.

## Citation

The ORACLE authors ask that any publication using their data cite:

```bibtex
@inproceedings{sankhe2019oracle,
  title={ORACLE: Optimized Radio clAssification through Convolutional neuraL nEtworks},
  author={Sankhe, Kunal and Belgiovine, Mauro and Zhou, Fan and Riyaz, Shamnaz and Ioannidis, Stratis and Chowdhury, Kaushik},
  booktitle={IEEE INFOCOM 2019 - IEEE Conference on Computer Communications},
  pages={370--378},
  year={2019}
}
```