M64/sid-music

---
license: mit
task_categories:
- text-generation
language:
- en
tags:
- music
- audio
- commodore-64
- sid
- chiptune
size_categories:
- 100M<n<1B
---

# SID Music Dataset

Register dumps from 2,418 Commodore 64 SID files for training music generation models.
9000 frames for each file, corresponding to 3 minutes of the sid file.

## Dataset Description

- **Source**: [HVSC](https://hvsc.c64.org/) (High Voltage SID Collection)
- **Size**: 1GB of register dump sequences
- **Format**: Hex-encoded SID register states at 50Hz
- **Songs**: 2,410 files from 15 legendary composers

## Composers Included

| Composer | Songs |
|----------|-------|
| DRAX (Thomas Mogensen) | 1042 |
| Laxity (Thomas E. Petersen) | 274 |
| Jeroen Tel | 163 |
| Thomas Detert | 162 |
| Reyn Ouwehand | 124 |
| David Whittaker | 98 |
| Ben Daglish | 86 |
| Johannes Bjerregaard | 84 |
| Rob Hubbard | 78 |
| Jonathan Dunn | 67 |
| Matt Gray | 47 |
| Charles Deenen | 46 |
| Chris Hülsbeck | 42 |
| Mark Cooksey | 39 |
| Martin Galway | 34 |
| **Total** | **2,418** |

## Data Format

Each frame is 25 SID registers encoded as 50 hex characters:
```
B0080005410A306011C0064108200016800D41082000B4031F
B0084005410A30601100074108200016C00D41082000B4031F
B0088005410A30601140074108200016000E41082000B4031F
...
<end>
```

- 50 hex characters = 25 bytes (SID registers $D400-$D418)
- `<end>` marks song boundaries
- 50 frames = 1 second of audio

## Register Layout
```
Bytes 0-6: Voice 1 (freq, pulse width, control, envelope)
Bytes 7-13: Voice 2
Bytes 14-20: Voice 3
Bytes 21-24: Filter + Volume
```
## Usage

### Quick Start with SidGPT
```bash
# Clone SidGPT
git clone https://github.com/M64GitHub/SidGPT
cd SidGPT
pip install torch numpy tqdm

# Download this dataset
wget https://huggingface.co/datasets/M64/sid-music/resolve/main/training.txt.gz
gunzip training.txt.gz
mv training.txt training/data/sid/input.txt

# Tokenize & Train
cd training/data/sid && python prepare.py && cd ../..
python train.py config/train_sid.py
```

### Or Use Pre-trained Model

Skip training entirely:
- [SID-GPT 25M Model](https://huggingface.co/M64/sid-gpt-25m)

### Manual / Custom Training

If using your own training setup:

1. **Download**: `training.txt.gz` (~100MB compressed, ~1GB uncompressed)
2. **Format**: Character-level, 22-token vocabulary
3. **Tokenize**: Map characters to indices:
```python
vocab = ['\n', '0', '1', '2', '3', '4', '5', '6', '7', '8', '9',
'A', 'B', 'C', 'D', 'E', 'F', '<', '>', 'd', 'e', 'n']
char_to_idx = {c: i for i, c in enumerate(vocab)}
```
4. **Train**: Any GPT/transformer architecture works. Recommended:
- Block size: 1020+ tokens (20+ frames context)
- Character-level prediction (no BPE)

### Pre-trained Model

Skip training and use the pre-trained model directly:
- [SID-GPT 25M](https://huggingface.co/M64/sid-gpt-25m)

### Statistics

- Total characters: ~1,000,000,000
- Vocabulary: 22 tokens (`0-9`, `A-F`, `<`, `>`, `d`, `e`, `n`, `\n`)
- Average song length: 9000 frames (~ 3 minutes)

## License

MIT License.

Original SID files from HVSC are © their respective composers.
This dataset contains derived register dumps for research purposes.

## Citation
```bibtex
@misc{sidmusicdataset2026,
author = {Mario Schallner},
title = {SID Music Dataset: C64 Register Dumps for ML},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/M64/sid-music}
}
```

## Related

- [SID-GPT Model](https://huggingface.co/M64/sid-gpt-25m)
- [SidGPT GitHub](https://github.com/M64GitHub/SidGPT)
- [HVSC](https://hvsc.c64.org/)