atoof/fma-music-descriptions

🎵 Free Music Archive with Full Music Flamingo Descriptions

A curated collection of 594 high-quality music tracks from the Free Music Archive, with complete semantic descriptions generated by NVIDIA's Music Flamingo model.

✨ What's New

This dataset includes the full Music Flamingo descriptions, not just extracted tags. Each track has:

• —📝 Complete textual description (mood, energy, instrumentation, production, use cases)
• —🏷️ Extracted semantic tags
• —🎵 High-quality audio (MP3)
• —📜 Full attribution and licensing

📊 Dataset Summary

• —Total Tracks: 594 music tracks
• —Total Duration: ~20 hours of audio
• —Format: MP3, embedded in Parquet
• —Descriptions: Full Music Flamingo semantic descriptions
• —Tags: 275 unique descriptive tags
• —License: CC BY 4.0

🎯 Use Cases

• —Audio Captioning: Train models to generate descriptions of music
• —Music Generation: Condition generation models on detailed descriptions
• —Audio-to-Text: Learn semantic understanding of music
• —Text-to-Music: Use descriptions as training data for text-to-music models
• —Music Information Retrieval: Semantic search and analysis
• —Audio Classification: Tag prediction from audio

📁 Dataset Structure

Features

Example Description

- Mood: tense, ominous, suspenseful, claustrophobic, mysterious, unsettling, dramatic.
- Energy: low, slow, hypnotic, minimalistic.
- Instrumentation & production: sparse, deep sub‑bass, metallic textures, atmospheric drones,
  resonant percussion (bottles/glasses), reverberant space, underscore/background music style.
- The track creates a dark, cinematic atmosphere perfect for suspenseful scenes, psychological
  thrillers, or mysterious documentary sequences. The acoustic elements (bottles and glasses)
  are heavily processed to create an otherworldly, tension-filled soundscape.

💻 Usage

Load the Dataset

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("parquet", data_files="train.parquet")

# Access an example
example = dataset["train"][0]
print("Title:", example["title"])
print("Description:", example["description"])
print("Tags:", example["tags"])

# Access audio
audio_array = example["audio"]["array"]
sample_rate = example["audio"]["sampling_rate"]

Use Descriptions for Music Generation

from audiocraft.models import MusicGen

# Use full descriptions to generate similar music
model = MusicGen.get_pretrained('melody')

# Take a description from the dataset
description = dataset["train"][0]["description"]
wav = model.generate([description])

Train Audio Captioning Model

# Use this dataset to train models that generate descriptions from audio

for example in dataset["train"]:
    audio = example["audio"]["array"]
    description = example["description"]  # Ground truth description
    tags = example["tags"]  # Alternative: use tags

    # Train your model here

Semantic Music Search

from sentence_transformers import SentenceTransformer
from sklearn.metrics.pairwise import cosine_similarity

# Encode all descriptions
model = SentenceTransformer('all-MiniLM-L6-v2')
descriptions = [ex["description"] for ex in dataset["train"]]
embeddings = model.encode(descriptions)

# Search with natural language
query = "dark cinematic orchestral music with tension and drama"
query_emb = model.encode([query])
similarities = cosine_similarity(query_emb, embeddings)

# Get top matches
top_idx = similarities[0].argsort()[-5:][::-1]
for idx in top_idx:
    print(f"{dataset['train'][idx]['title']} by {dataset['train'][idx]['artist']}")

📈 Dataset Statistics

• —Total Tracks: 594
• —Unique Artists: 44
• —Total Tags: 5,962 tag instances
• —Unique Tags: 275
• —Average Tags per Track: ~9.3
• —Average Description Length: ~500 characters
• —Total Duration: ~20 hours

Top Tags

1. 1.dark (412 tracks)
2. 2.cinematic (363 tracks)
3. 3.tense (292 tracks)
4. 4.ambient (268 tracks)
5. 5.suspenseful (258 tracks)

🔬 Description Generation

The full descriptions were generated using NVIDIA's Music Flamingo model, which provides:

• —Mood analysis: Emotional characteristics and atmosphere
• —Energy assessment: Tempo, intensity, and pacing
• —Instrumentation: Detailed breakdown of instruments and sounds
• —Production style: Mixing, effects, and sonic characteristics
• —Use case suggestions: Potential applications for the music

📜 License & Attribution

This dataset contains music from the Free Music Archive, licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).

Attribution Required

When using this dataset:

1. 1.Provide attribution to original artists (in attribution field)
2. 2.Link back to Free Music Archive
3. 3.Credit NVIDIA Music Flamingo for descriptions
4. 4.Indicate any modifications made

🙏 Acknowledgments

• —[Free Music Archive](https://freemusicarchive.org/): High-quality, openly licensed music
• —Artists: All 44 artists who contributed their music
• —[NVIDIA Music Flamingo](https://huggingface.co/nvidia/music-flamingo-2601-hf): State-of-the-art music understanding model
• —Dataset Creator: Dataset curation and processing

📝 Citation

@dataset{fma_music_flamingo_2026,
  title={Free Music Archive with Full Music Flamingo Descriptions},
  author={Free Music Archive and NVIDIA},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/atoof/fma-music-descriptions}}
}

Please also cite Music Flamingo:

@misc{music-flamingo-2601,
  title={Music Flamingo: A Large-Scale Music Understanding Model},
  author={NVIDIA},
  year={2026},
  howpublished={\url{https://huggingface.co/nvidia/music-flamingo-2601-hf}}
}

🔗 Links

• —Free Music Archive
• —NVIDIA Music Flamingo
• —Original FMA Dataset

Perfect for: Audio captioning research, music generation training, semantic music understanding, audio-to-text models