README.md

---
title: LEMM Training Data Repository
emoji: 🎵
colorFrom: purple
colorTo: pink
sdk: static
pinned: false
license: mit
---

# 🎵 LEMM Training Data Repository

Central storage for all LEMM training artifacts:
- **LoRA Adapters** - User-trained model adaptations
- **Training Datasets** - Prepared and curated audio datasets

Part of the [LEMM (Let Everyone Make Music)](https://huggingface.co/spaces/Gamahea/lemm-test-100) project.

---

## 📁 Repository Structure

```
lemmdata/
├── loras/                          # LoRA Adapters (ZIP files)
│   ├── jazz-v1.zip                # Example: Jazz style adapter
│   ├── metal-data1.zip            # Example: Metal style adapter
│   └── ...
│
└── datasets/                       # Training Datasets (ZIP files)
    ├── gtzan_prepared.zip         # Example: GTZAN dataset
    ├── user_dataset_123.zip       # Example: User-uploaded dataset
    └── ...
```

---

## 🎨 LoRA Adapters (`loras/`)

Each LoRA is packaged as a ZIP file containing:

- **`final_model.pt`** - Trained LoRA weights (PyTorch checkpoint)
- **`config.yaml`** - Training hyperparameters and settings
- **`metadata.json`** - Training statistics, timestamps, dataset info
- **`README.md`** - Documentation and usage instructions

### How to Use

**Download in LEMM:**
1. Go to [LEMM Space](https://huggingface.co/spaces/Gamahea/lemm-test-100)
2. Navigate to "LoRA Management" tab
3. Click "Sync from HuggingFace"
4. Select LoRA from dropdown
5. Use in generation or continue training

**Download via Code:**
```python
from huggingface_hub import hf_hub_download
import zipfile

# Download LoRA ZIP
zip_path = hf_hub_download(
    repo_id="Gamahea/lemmdata",
    repo_type="dataset",
    filename="loras/jazz-v1.zip"
)

# Extract
with zipfile.ZipFile(zip_path, 'r') as zipf:
    zipf.extractall("./my_loras/jazz-v1")
```

---

## 📊 Training Datasets (`datasets/`)

Each dataset is packaged as a ZIP file containing:

- **`dataset_info.json`** - Metadata (size, format, split ratios)
- **`train/`** - Training audio files
- **`val/`** - Validation audio files

### Supported Formats

- **Audio**: WAV, MP3, FLAC, OGG
- **Sample Rate**: 44.1kHz or 48kHz recommended
- **Channels**: Mono or Stereo

### How to Use

**Download in LEMM:**
1. Go to [LEMM Space](https://huggingface.co/spaces/Gamahea/lemm-test-100)
2. Navigate to "Training" tab
3. Click "Import Dataset" → "From HuggingFace"
4. Select dataset
5. Use for training

**Download via Code:**
```python
from huggingface_hub import hf_hub_download
import zipfile

# Download dataset ZIP
zip_path = hf_hub_download(
    repo_id="Gamahea/lemmdata",
    repo_type="dataset",
    filename="datasets/gtzan_prepared.zip"
)

# Extract
with zipfile.ZipFile(zip_path, 'r') as zipf:
    zipf.extractall("./my_datasets/gtzan_prepared")
```

---

## 🚀 Contributing

### Upload LoRA

Train a LoRA in [LEMM Space](https://huggingface.co/spaces/Gamahea/lemm-test-100):
1. Prepare or select a dataset
2. Configure training parameters
3. Start training
4. LoRA automatically packaged as ZIP and uploaded to `loras/{your-lora-name}.zip`

### Upload Dataset

Prepare a dataset and export:
1. Upload audio files to LEMM
2. Use dataset preparation tools
3. Export as prepared dataset
4. Dataset packaged as ZIP and uploaded to `datasets/{your-dataset-name}.zip`

---

## 📝 Naming Conventions

### LoRA Names
- **Format**: `{style}-{variant}_{version}`
- **Examples**: 
  - `jazz-bebop_v1`
  - `rock-heavy_v2`
  - `classical-piano_v1`

### Dataset Names
- **Format**: `{source}_{description}`
- **Examples**:
  - `gtzan_prepared`
  - `user_dataset_1702987654`
  - `opensinger_vocals`

---

## 🔐 Authentication

**Read Access**: Public (no authentication required)

**Write Access**: Requires HuggingFace token
- Only LEMM Space can upload
- User-trained artifacts auto-upload
- Token managed via HF Space secrets

---

## 📚 Related Resources

- **LEMM Space**: https://huggingface.co/spaces/Gamahea/lemm-test-100
- **GitHub Repo**: https://github.com/Gamahea/Angen
- **DiffRhythm2**: Music generation model with vocals
- **MuQ-MuLan**: Music style encoding

---

## 🎓 Training Best Practices

### LoRA Training
- **Dataset Size**: 100+ clips minimum
- **LoRA Rank**: 8-32 for most styles
- **Learning Rate**: 1e-4 to 1e-3
- **Epochs**: 20-50 depending on dataset

### Dataset Preparation
- **Clip Length**: 10-30 seconds
- **Audio Quality**: Clean, well-produced
- **Consistency**: Similar genre/style
- **Diversity**: Varied within target style

---

## 📄 License

MIT License - Free to use, modify, and share.

All contributed LoRAs and datasets inherit this license unless otherwise specified.

---

## 🏷️ Tags

`music-generation` `lora` `diffrhythm2` `audio` `training-data` `datasets` `models` `lemm`

---

**Last Updated**: December 2025  
**Repository**: https://huggingface.co/datasets/Gamahea/lemmdata  
**LEMM Space**: https://huggingface.co/spaces/Gamahea/lemm-test-100