Datasets:

cpnlab
/

LSM-Tokenized-Full

The dataset is designed to bridge wireless signal processing and large language models (LLMs) by converting raw spectrum measurements into discrete token sequences, enabling direct use with transformer-based architectures.

🚀 Key Highlights

📊 Real-world large-scale dataset
- ~21 TB raw RF data
- 33 sub-GHz frequency bands (54 MHz – 990 MHz)
- ~8.4 billion tokens
🧠 LLM-ready format
- Fully tokenized sequences
- Compatible with GPT-style models, LLaMA, Mistral, etc.
🔬 Rich metadata integration
- Gain, frequency, time, and spectrum structure embedded as tokens
⚡ Efficient representation
- Vocabulary size: 128 tokens
- Stored as 1-byte integers → fast I/O and low memory footprint

📁 Dataset Structure

Each sample is a fixed-length token sequence:

[ METADATA TOKENS (36) | SPECTRUM TOKENS (256) ]

Total sequence length: 292 tokens
Input / Target split:
- Input: first 128 spectrum tokens
- Target: next 128 tokens (forecasting task)

🔑 Tokenization Scheme

1. Spectrum (PSD)

Range: -118 dBm to -18 dBm
Quantized to 101 discrete levels
Token IDs: 1 – 101

2. Metadata Tokens

Feature	Description
Frequency	33 bands
Gain	0–79 dB
Frequency Bin	Encoded using 2 tokens (base-16)
Timestamp	Decomposed into:
	Day, Date, Month, Year, Hour, Minute, Second

3. Special Tokens

0 → Padding
102–127 → Structural tags (start/end markers for fields)

🧬 Data Generation Pipeline

Raw IQ Data
   ↓
STFT (Spectrogram)
   ↓
Max-Pooling (temporal)
   ↓
Trimmed Mean (robust smoothing)
   ↓
256 × 256 Spectrogram
   ↓
Tokenization

Key properties:

Time resolution: ~3.91 ms
Frequency resolution: 78.125 kHz
Designed for long-sequence modeling

🎯 Task Definition

Spectrum Forecasting

Given:

Past PSD tokens (t = 0 → 127)

Predict:

Future PSD tokens (t = 128 → 255)

This formulation allows:

Autoregressive modeling (GPT-style)
Sequence-to-sequence learning
Diffusion or masked modeling approaches

📊 Dataset Characteristics

Highly diverse spectrum environments
Includes high-variance bands (e.g., 630 MHz, 650 MHz) for robustness evaluation
Captures:
- Cellular activity
- Broadcast signals
- Environmental noise patterns

🔧 Usage

Load Dataset (Hugging Face)

from datasets import load_dataset

dataset = load_dataset("cpnlab/LSM-Tokenized-Full")
sample = dataset["train"][0]

tokens = sample["input_ids"]

Typical Training Setup

Vocabulary size: 128
Input length: 164 (metadata + partial sequence)
Output length: 128
Loss:
- Cross-Entropy
- Optional RMSE hybrid (for high-variance bands)

🧠 Compatible Models

This dataset is designed for:

Transformer-based models:
- GPT-style (decoder-only)
- LLaMA / Mistral / Gemma / Phi variants
Sequence models:
- LSTM / RNN baselines
Generative approaches:
- Diffusion models
- Masked modeling

📦 Related Resources

💻 Code: https://github.com/UNL-CPN-Lab/LSM
🌐 Project Page: https://lsm.unl.edu
📄 Paper: DySPAN 2026 (see above)

⚠️ Notes

This dataset is tokenized — raw IQ data is not included here
Designed for research in AI-driven spectrum intelligence
Suitable for:
- Dynamic Spectrum Access (DSA)
- 6G wireless systems
- RF foundation models

📜 Citation

If you use this dataset, please cite:

@inproceedings{lunar2026lsm,
  title={Large Spectrum Models (LSMs): Decoder-Only Transformer-Powered Spectrum Activity Forecasting via Tokenized RF Data},
  author={Lunar, Mohammad Mosiur and Vuran, Mehmet C.},
  booktitle={IEEE DySPAN},
  year={2026}
}

🤝 Acknowledgment

Collected using the NEXTT city-scale wireless testbed and supported by NSF grants.

🔥 Why This Dataset Matters

This is one of the first datasets that:

Treats RF signals as language tokens
Enables foundation models for wireless spectrum
Scales to billions of tokens in real-world environments