---
language:
- en
task_categories:
- text-to-audio
- audio-to-audio
tags:
- music
- midi
- chroma
- music-generation
- geometric-deep-learning
size_categories:
- 1K<n<10K
---

# MIDI Chroma Dataset

This version requires the genres to be fixed and a restructure before it's compatible with training.


Pre-processed version of [foldl/midi](https://huggingface.co/datasets/foldl/midi) with chroma features extracted directly from MIDI note events.

## Dataset Description

This dataset contains **4719 songs** with pre-computed chroma features for efficient music generation training.

### Features

- **name**: Song title (string)
- **genre**: List of genres (list of strings)
- **chroma**: Pre-computed chroma features `[128, 12]` (float32 array)
  - 12 pitch classes (C, C#, D, D#, E, F, F#, G, G#, A, A#, B)
  - 128 time steps
  - Values normalized to sum to 1.0 per timestep
- **text**: Text description for conditioning (string)

### Extraction Method

Chroma features are extracted **directly from MIDI note events** without audio synthesis:
- Notes are mapped to their pitch class (0-11)
- Velocity is used for intensity weighting
- Temporal resolution: ~10 FPS
- Much faster than audio-based extraction

### Usage
```python
from datasets import load_dataset
import torch

# Load dataset
dataset = load_dataset("AbstractPhil/foldl-midi")

# Access samples
sample = dataset['train'][0]
chroma = torch.tensor(sample['chroma'])  # [128, 12]
text = sample['text']                     # "rock, pop: Genesis - The Light Dies Down"

print(f"Text: {text}")
print(f"Chroma shape: {chroma.shape}")
```

### Training ChromaLyra

This dataset is designed for training **ChromaLyra**, a geometric VAE for music generation:
```python
from geovocab2.train.model.chroma.chroma_lyra import ChromaLyra, ChromaLyraConfig

config = ChromaLyraConfig(
    n_chroma=12,
    seq_len=128,
    latent_dim=256,
    hidden_dim=384
)

model = ChromaLyra(config)
# Train with text conditioning...
```

## Dataset Creation

Created by extracting chroma from valid MIDI files in foldl/midi dataset:
- Filtered songs: 1s - 3min duration
- Skipped empty/drum-only tracks
- Original: ~20K MIDI files → This dataset: ~4719 valid samples

## Citation

Original dataset:
```bibtex
@misc{foldl-midi,
  author = {foldl},
  title = {MIDI Dataset},
  year = {2023},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/foldl/midi}
}
```

Geometric approach:
```bibtex
@misc{abstract-phil-geovocab,
  author = {AbstractPhil},
  title = {GeoVocab: Geometric Deep Learning for Music Generation},
  year = {2025},
  url = {https://github.com/AbstractPhil/geovocab2}
}
```

## License

Same as original foldl/midi dataset; 
https://huggingface.co/datasets/foldl/midi

## Acknowledgments

- Original MIDI dataset: foldl
- Chroma extraction: pretty_midi library
- Geometric VAE architecture: AbstractPhil/GeoVocab2