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

Feature	Type	Description
audio	Audio	Audio data (MP3 format, various sample rates)
title	string	Track title
artist	string	Artist name
description	string	Full Music Flamingo description (mood, energy, instrumentation, etc.)
tags	list[string]	Extracted semantic tags from description
genre_tags	list[string]	Original genre classifications from FMA
track_id	string	Free Music Archive track ID
original_page	string	Link to original FMA page
attribution	string	License and attribution information

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. dark (412 tracks)
2. cinematic (363 tracks)
3. tense (292 tracks)
4. ambient (268 tracks)
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. Provide attribution to original artists (in attribution field)
2. Link back to Free Music Archive
3. Credit NVIDIA Music Flamingo for descriptions
4. Indicate any modifications made

🙏 Acknowledgments

• Free Music Archive: High-quality, openly licensed music
• Artists: All 44 artists who contributed their music
• NVIDIA Music Flamingo: 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