LyricGen Dataset

A curated dataset of musical features paired with song lyrics for training language models to generate lyrics based on audio characteristics.

Dataset Description

The LyricGen dataset contains 605 musical feature-lyric pairs designed for fine-tuning large language models (like Llama 3.2) to generate song lyrics conditioned on musical characteristics. Each entry provides structured musical features extracted from audio alongside corresponding song lyrics.

Dataset Structure

The dataset is provided in JSONL format with the following structure:

{
  "input": "Generate lyrics for a song with these characteristics:\nTempo: 128.3 BPM\nDuration: 2:36\nLoudness: -14.3 dB\nKey: F minor\nTime Signature: 1/4\n\nLyrics:",
  "output": "[Song lyrics content]",
  "full_prompt": "[Complete formatted prompt for training]"
}

Fields

• input: Formatted prompt containing musical features and instruction
• output: Corresponding song lyrics
• full_prompt: Complete training prompt in chat format

Musical Features

Each entry includes five key musical characteristics:

1. Tempo: Song tempo in beats per minute (BPM)
2. Duration: Song length in MM:SS format
3. Loudness: Audio loudness level in decibels (dB)
4. Key: Musical key (e.g., C major, F minor)
5. Time Signature: Rhythmic structure (e.g., 4/4, 3/4)

Data Creation Process

Source Data

• Primary Source: Million Song Dataset
• Lyrics Source: Genius API
• Initial Sample: 2,000 songs from Million Song Dataset
• Final Dataset Size: 605 songs (30.25% retention rate)

Processing Pipeline

1. Feature Extraction: Five musical features were extracted from the Million Song Dataset
2. Lyrics Retrieval: Song lyrics were obtained using the Genius API by matching song names
3. Data Filtering:
   • Removed songs without available lyrics
   • Filtered to ensure musical features align with librosa library standards
   • Quality control for feature accuracy and completeness
4. Prompt Formatting: Data structured for instruction-following training compatible with models like Llama 3.2

Quality Assurance

• Musical features validated against librosa library specifications
• Lyrics matched and verified through Genius API
• Prompt formatting optimized for language model training

Intended Use

Primary Use Case

Fine-tuning large language models to generate song lyrics based on musical characteristics.

Compatible Models

• Llama 3.2 8B Instruct
• Other instruction-following language models
• Models supporting chat/conversation formats

Applications

• Music composition assistance
• Lyric generation tools
• Creative writing applications
• Music production workflows

Dataset Statistics

• Total Examples: 605
• Source Retention Rate: 30.25% (605/2000)
• Musical Features: 5 per song
• Format: JSONL
• Split: Training and validation sets available

Usage Example

import json
from datasets import Dataset

# Load the dataset
def load_lyricgen_data(file_path):
    data = []
    with open(file_path, 'r', encoding='utf-8') as f:
        for line in f:
            data.append(json.loads(line.strip()))
    return data

# Load and use
train_data = load_lyricgen_data('train.jsonl')
dataset = Dataset.from_list(train_data)

# Example entry
print("Musical Features Input:", train_data[0]['input'])
print("Lyric Output:", train_data[0]['output'][:100] + "...")

Limitations and Considerations

• Copyright: Lyrics are sourced from existing songs and may be subject to copyright restrictions
• Bias: Dataset reflects biases present in the Million Song Dataset and available lyrics on Genius
• Coverage: Limited to songs with available lyrics on Genius API
• Language: Primarily English lyrics
• Musical Scope: Features limited to five basic characteristics

Citation

If you use this dataset in your research, please cite:

@dataset{lyricgen2024,
  title={LyricGen: A Musical Features and Lyrics Dataset for Language Model Training},
  author={umerbappi},
  year={2024},
  url={https://huggingface.co/datasets/umerbappi/LyricGen}
}

Data Sources

• Million Song Dataset: Musical feature extraction
• Genius API: Lyrics retrieval and matching
• Librosa Library: Musical feature standardization

License

Please note that while the dataset structure and processing methodology are provided, individual song lyrics may be subject to copyright restrictions. Users should ensure compliance with applicable copyright laws when using this dataset.

Contact

For questions about the dataset or requests for additional information, please contact the dataset creator through the Hugging Face dataset repository.