JamendoMaxCaps Dataset

JamendoMaxCaps is a large-scale dataset of over 362,000 instrumental tracks sourced from the Jamendo platform. It includes generated music captions and original metadata. Additionally, we introduce a retrieval system that utilizes both musical features and metadata to identify similar songs, which are then used to impute missing metadata via a local large language model (LLLM).

This dataset facilitates research in:

• Music-language understanding
• Music retrieval and representation learning
• AI-generated music applications

For details on the retrieval system and metadata imputation, visit our GitHub repository.

Dataset Details

The dataset is stored in parquet format within the data folder. Each entry contains:

• array: The audio file stored as an array.
• path: The unique identifier for the MP3 file (<unique_id>.mp3).

Extracting MP3 Files

To convert the stored audio arrays into MP3 files, use the following script:

import os
import soundfile as sf
from datasets import load_dataset

def read_jamendo_max():
    dataset = load_dataset("amaai-lab/JamendoMaxCaps", data_dir="data")
    return dataset

def save_mp3_files(dataset, output_dir="mp3_files"):
    os.makedirs(output_dir, exist_ok=True)
    
    for example in dataset['train']:
        audio = example['audio']['array']
        sample_rate = example['audio']['sampling_rate']
        path = example['audio']['path']
        output_path = os.path.join(output_dir, path)
        
        sf.write(output_path, audio, sample_rate, format='MP3')
        print(f"Saved file: {output_path}")

if __name__ == "__main__":
    dataset = read_jamendo_max()
    save_mp3_files(dataset)

Metadata

Original metadata is provided in multiple JSONL files for easy handling. Each entry follows this structure:

{
  "id": "882977",
  "name": "A Singleness of Intention 04",
  "duration": 154,
  "artist_id": "362132",
  "artist_name": "Matthew VandenBrook",
  "album_name": "A Singleness of Intention",
  "album_id": "103795",
  "license_ccurl": "http://creativecommons.org/licenses/by-nc-nd/3.0/",
  "releasedate": "2011-12-18",
  "audio": "https://prod-1.storage.jamendo.com/?trackid=882977&format=mp31",
  "audiodownload": "https://prod-1.storage.jamendo.com/download/track/882977/mp32/",
  "musicinfo": {
    "vocalinstrumental": "instrumental",
    "speed": "medium",
    "tags": {
      "genres": [],
      "instruments": [],
      "vartags": []
    }
  },
  "audiodownload_allowed": true
}

• id: Unique identifier for each song.
• musicinfo: Contains attributes like instrumental/vocal classification and tempo (e.g., medium speed).
• audio & audiodownload: Links to the track's online and downloadable versions.

Captions

We provide captions for 30-second segments using Qwen2-Audio. Captions are stored in final_caption30sec.jsonl. Example:

{
  "start_time": 60.0,
  "end_time": 90.0,
  "caption": "The music is a fast-paced electronic track in F# minor with a tempo of 84 BPM featuring heavy drum beats and a complex chord progression. The genre includes breakbeat, drum and bass, easy listening, electronic, IDM, and lounge. It creates an energetic mood and could be suitable for a mobile car racing game soundtrack.",
  "id": "1176956"
}

Links

• Code Repository: GitHub
• Paper: arXiv

Citation

If you use this dataset, please cite:

@article{roy2025jamendomaxcaps,
  title={JamendoMaxCaps: A Large Scale Music-caption Dataset with Imputed Metadata},
  author={Roy, Abhinaba and Liu, Renhang and Lu, Tongyu and Herremans, Dorien},
  journal={arXiv:2502.07461},
  year={2025}
}

License

License: CC-BY-SA 3.0