---
license: mit
---

## Citation
If you use this dataset, please cite:

1) [A Multi-Task Foundation Model for Wireless Channel Representation Using Contrastive and Masked Autoencoder Learning](https://arxiv.org/abs/2505.09160) 
```bibtex
@article{guler2025multitask,
  title={A Multi-Task Foundation Model for Wireless Channel Representation Using Contrastive and Masked Autoencoder Learning},
  author={Guler, Berkay and Geraci, Giovanni and Jafarkhani, Hamid},
  journal={arXiv preprint arXiv:2505.09160},
  year={2025},
  url={https://arxiv.org/abs/2505.09160}
}
```

See also our earlier workshop paper that introduced this dataset:

2) [Robust Channel Representation for Wireless: A Multi-Task Masked Contrastive Approach](https://openreview.net/forum?id=KXNDs9ZGb9)

```bibtex
@inproceedings{guler2025robust,
  title={Robust Channel Representation for Wireless: A Multi-Task Masked Contrastive Approach},
  author={Guler, Berkay and Geraci, Giovanni and Jafarkhani, Hamid},
  booktitle={NeurIPS 2025 Workshop on AI for Next Generation Wireless (AI4NextG)},
  year={2025},
  url={https://openreview.net/forum?id=KXNDs9ZGb9}
}
```

## Loading the Data

```python
from pathlib import Path
import numpy as np
from tqdm import tqdm

def get_sample_file_content(file_path):
    
    def get_key_data_pairs(file_path, allow_empty=False):
        data_dict = {}
        with np.load(file_path, allow_pickle=True) as file_content:
            keys = file_content.keys()
            for key in keys:
                if key not in data_dict:
                    if not allow_empty and len(file_content[key].shape) == 0:
                        continue
                    data_dict[key] = file_content[key] 
        return data_dict
    
    for file in tqdm(file_path.iterdir(), desc="Loading pretraining files"):
        if file.is_file() and not file.name.startswith("."):
            return get_key_data_pairs(file)
            
pretrain_path = Path.cwd() / "sub6" / "pretrain"

sample_data = get_sample_file_content(pretrain_path)
for key, data in sample_data.items():
    print(f"Key {key} has shape:\t{data.shape}")
```

    Loading pretraining files: 0it [00:00, ?it/s]

    Key channels has shape:	(10451, 1, 32, 32)
    Key rx_pos has shape:	(10451, 3)
    Key tx_pos has shape:	(1, 3)
    Key los has shape:	(10451,)
    Key active_mask_original_indices has shape:	(10451,)
    Key beam_labels has shape:	(10451, 3)

- `channels` stores the channels with shape `(number of channels, number of rx antennas, number of tx antennas, number of subcarriers)`
- `rx_pos` stores the 3D positions of the users
- `tx_pos` stores the TX position
- `los` stores a binary array of `0`'s and `1`'s corresponding to nLoS and LoS channels, respectively.
- `beam_labels[:0]` stores the best beam index from a uniform codebook with size 16
- `beam_labels[:1]` stores the best beam index from a uniform codebook with size 32
- `beam_labels[:2]` stores the best beam index from a uniform codebook with size 32

## Folder Structure
```
├── sub 6/  # Contains 3.5 GHz channel samples
│   ├── pretrain/  # Used for pretraining
│   │   ├── boston5g_3p5_bs000.npz
│   │   ├── city_66_bruxelles_3p5_bs000.npz
│   │   └── ...
│   └── task/  # Used for downstream task evaluations
│       ├── train/  # Can be used if finetuning is performed
│       │   ├── city_7_sandiego_3p5_bs000.npz
│       │   ├── city_7_sandiego_3p5_bs001.npz
│       │   └── ...
│       └── test/  # Held-out test files for downstream tasks
│           ├── city_3_houston_3p5_bs000.npz
│           ├── city_3_houston_3p5_bs001.npz
│           └── ...
│
└── mmwave/  # Contains 28 GHz channel samples
    └── task/  # Used for downstream task evaluations
        ├── train/  # Can be used if finetuning is performed
        │   ├── city_7_sandiego_28_bs000.npz
        │   ├── city_7_sandiego_28_bs001.npz
        │   └── ...
        └── test/  # Held-out test files for downstream tasks
            ├── city_3_houston_28_bs000.npz
            ├── city_3_houston_28_bs001.npz
            └── ...
```
## Dataset Properties

| Parameter | Value |
|-----------|-------|
| Num. Subcarriers | 32 |
| Subcarrier spacing | 30 kHz |
| Total bandwidth | 960 kHz |
| TX Antennas | 32 (half-wavelength spacing) |
| Radiation pattern | Isotropic, single polarization |
| Maximum propagation paths | 20 |

## Dataset Generation

Using the [DeepMIMO](https://www.deepmimo.net/) tool, we generate **2.5M samples** at **3.5 GHz** operating frequency from **56 scenarios** for pretraining. Each sample corresponds to a channel for a distinct user-base station pair from one of the following scenarios:

**Pretraining Scenarios (56 cities):**
Amsterdam, ASU Campus, Athens, Bangkok, Barcelona, Beijing, Boston, Brussels, Cairo, Cape Town, Charlotte, Chicago, Denver, Dubai, Edinburgh, Florence, Fort Worth, Fujiyoshida, Granada, Hatsukaichi, Helsinki, Hong Kong, Indianapolis, Istanbul, Jerusalem, Kyoto, Havana, Lisbon, Los Angeles, Madrid, Marrakesh, Mumbai, New Delhi, New York, North Jakarta, Oklahoma City, Philadelphia, Phoenix, Reykjavik, Rio de Janeiro, Rome, San Francisco, San Nicolas, Saint Petersburg, Santa Clara, Santiago, Seattle, Seoul, Singapore, Stockholm, Sumida City, Sydney, Taipei, Taito City, Toronto, and Warsaw.

## Downstream Task Dataset

We generate a separate dataset for downstream tasks with **0.55M total samples** from **10 unseen scenarios**:

**Downstream Scenarios (10 cities):**
Austin, Centro, Columbus, Dallas, Gurbchen, Houston, Miami, Montreal, Prague, and San Diego.

Similar to the pretraining dataset, some scenarios feature more than one base station with uniformly distributed users.

## Additional Configurations
- **28 GHz channels** are generated for downstream task scenarios with available ray-tracing data