Archit00/jamendo-qa-mirror

Jamendo-QA: A Large-Scale Music Question Answering Dataset

Jamendo-QA is a large-scale dataset designed for music-related question answering (Music-QA) research.
It is built upon the Jamendo Music collection and supports research in music knowledge QA, audio-text multimodal learning, and music information retrieval (MIR).

---

📊 Dataset Summary

Item	Description
Name	Jamendo-QA
Domain	Music, Question Answering, MIR
Languages	English
Tasks	Question Answering, RAG, Multimodal Music Understanding
Audio Files	7,335 unique music tracks
Format	Parquet (with embedded audio bytes) + JSON
License	Research-only

---

📁 Dataset Structure

Main Files

File	Size	Description
jamendo_qa_dataset.parquet	39 GB	Complete dataset with audio bytes embedded
metadata.json	9.6 MB	Track metadata (artist, genre, duration, lyrics, etc.)
qa_v1.json	15 MB	QA pairs version 1 (basic metadata-based QA)
qa_v2.json	29 MB	QA pairs version 2 (detailed music analysis with captions)

---

🔧 Features

Parquet Dataset (jamendo_qa_dataset.parquet)

Feature	Type	Description
file_name	string	Unique identifier for the audio file
audio_path	string	Relative path to the audio file
audio_bytes	binary	Raw audio data in WAV format (16kHz)
duration	float	Length of the audio in seconds
artist	string	Artist name
genre	string	Music genre
speed	string	Tempo category (slow/medium/fast/high)
gender	string	Vocal gender (male/female/instrumental)
language	string	Language of lyrics
lyric	string	Lyric content
conversation	string	JSON-formatted QA conversation pairs

Metadata JSON (metadata.json)

Feature	Type	Description
file_name	string	Audio file name
audio_path	string	Relative path (audio/xxx.wav)
duration	int	Duration in seconds
artist_name	string	Artist name
genre	string	Music genre
speed	string	Tempo category
gender	string	Vocal gender
language	string	Language code
lyric	string	Full lyrics text

QA v2 JSON (qa_v2.json)

Feature	Type	Description
audio_name	string	Audio identifier
audio_path	string	Relative path to audio
prompt	string	Captioning prompt used
caption	string	Detailed music analysis/description
qa_pairs	array	List of question-answer pairs

---

💡 Usage

Load with Hugging Face Datasets

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("m-a-a-p/Jamendo-QA")

# Access the first example
example = dataset['train'][0]
print(f"File: {example['file_name']}")
print(f"Artist: {example['artist']}")
print(f"Genre: {example['genre']}")
print(f"Duration: {example['duration']} seconds")

# Parse QA conversation
import json
conversation = json.loads(example['conversation'])
for turn in conversation:
    print(f"{turn['from']}: {turn['value']}")

Load Metadata JSON

import json

with open('data/metadata.json', 'r') as f:
    metadata = json.load(f)

print(f"Total tracks: {len(metadata)}")
for track in metadata[:3]:
    print(f"  - {track['file_name']}: {track['genre']} by {track['artist_name']}")

Load QA v2 (Detailed Analysis)

import json

with open('data/qa_v2.json', 'r') as f:
    qa_data = json.load(f)

# Example: Get caption and QA pairs for first track
track = qa_data[0]
print(f"Track: {track['audio_name']}")
print(f"Caption: {track['caption'][:200]}...")
print(f"QA Pairs: {len(track['qa_pairs'])}")
for qa in track['qa_pairs'][:3]:
    print(f"  Q: {qa['question']}")
    print(f"  A: {qa['answer']}")

---

📈 Dataset Statistics

Statistic	Value
Total Tracks	7,335
Total Duration	~400 hours
Unique Genres	35+
Unique Artists	7,000+
QA Pairs (v1)	29,340 (4 per track)
QA Pairs (v2)	58,680 (8 per track)
Total QA Pairs	88,020

---

🔍 Citation

If you use Jamendo-QA in your research, please cite:

@misc{koh2025jamendoqalargescalemusicquestion,
      title={Jamendo-QA: A Large-Scale Music Question Answering Dataset}, 
      author={Junyoung Koh and Soo Yong Kim and Yongwon Choi and Gyu Hyeong Choi},
      year={2025},
      eprint={2509.15662},
      archivePrefix={arXiv},
      primaryClass={cs.MM},
      url={https://arxiv.org/abs/2509.15662}, 
}

---

📝 License

Jamendo-QA is available for research purposes only.
Please check the Hugging Face dataset page for download instructions and usage guidelines.

---

🙏 Acknowledgments

This dataset is built upon the Jamendo Music collection.
We thank the Jamendo community for making their music available for research purposes.