---
dataset_info:
  features:
  - name: analysis_sample_rate
    dtype: int32
  - name: artist_7digitalid
    dtype: int32
  - name: artist_familiarity
    dtype: float64
  - name: artist_hotttnesss
    dtype: float64
  - name: artist_id
    dtype: string
  - name: artist_latitude
    dtype: float64
  - name: artist_location
    dtype: string
  - name: artist_longitude
    dtype: float64
  - name: artist_mbid
    dtype: string
  - name: artist_mbtags
    sequence: binary
  - name: artist_mbtags_count
    sequence: int64
  - name: artist_name
    dtype: string
  - name: artist_playmeid
    dtype: int32
  - name: artist_terms
    sequence: binary
  - name: artist_terms_freq
    sequence: float64
  - name: artist_terms_weight
    sequence: float64
  - name: audio_md5
    dtype: string
  - name: bars_confidence
    sequence: float64
  - name: bars_start
    sequence: float64
  - name: beats_confidence
    sequence: float64
  - name: beats_start
    sequence: float64
  - name: danceability
    dtype: float64
  - name: duration
    dtype: float64
  - name: end_of_fade_in
    dtype: float64
  - name: energy
    dtype: float64
  - name: key
    dtype: int32
  - name: key_confidence
    dtype: float64
  - name: loudness
    dtype: float64
  - name: mode
    dtype: int32
  - name: mode_confidence
    dtype: float64
  - name: num_songs
    dtype: int64
  - name: release
    dtype: string
  - name: release_7digitalid
    dtype: int32
  - name: sections_confidence
    sequence: float64
  - name: sections_start
    sequence: float64
  - name: segments_confidence
    sequence: float64
  - name: segments_loudness_max
    sequence: float64
  - name: segments_loudness_max_time
    sequence: float64
  - name: segments_loudness_start
    sequence: float64
  - name: segments_pitches
    sequence:
      sequence: float64
  - name: segments_start
    sequence: float64
  - name: segments_timbre
    sequence:
      sequence: float64
  - name: similar_artists
    sequence: binary
  - name: song_hotttnesss
    dtype: float64
  - name: song_id
    dtype: string
  - name: start_of_fade_out
    dtype: float64
  - name: tatums_confidence
    sequence: float64
  - name: tatums_start
    sequence: float64
  - name: tempo
    dtype: float64
  - name: time_signature
    dtype: int32
  - name: time_signature_confidence
    dtype: float64
  - name: title
    dtype: string
  - name: track_7digitalid
    dtype: int32
  - name: track_id
    dtype: string
  - name: year
    dtype: int32
  splits:
  - name: train
    num_bytes: 2365768621
    num_examples: 10000
  download_size: 1041881893
  dataset_size: 2365768621
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# Million Song Subset (Processed Version)

## Overview
This dataset is a structured extraction of the [Million Song Subset](http://millionsongdataset.com/pages/getting-dataset/#subset), derived from HDF5 files into a tabular format for easier accessibility and analysis.

## Source
- Original dataset: **Million Song Dataset** (LabROSA, Columbia University & The Echo Nest)
- Subset used: **Million Song Subset** (10,000 songs)
- URL: [http://millionsongdataset.com](http://millionsongdataset.com)

## Processing Steps
1. **Extraction**: Used `hdf5_getters.py` to retrieve all available fields.
2. **Parallel Processing**: Optimized extraction with `ProcessPoolExecutor` for speed.
3. **Conversion**: Structured into a Pandas DataFrame.
4. **Storage**: Saved as a Parquet file for efficient usage.

## Format
- **Columns**: Contains all available attributes from the original dataset, including artist metadata, song features, and audio analysis.
- **File Format**: Parquet (optimized for efficient querying & storage).

## Usage
- Load the dataset with Datasets:
  ```python
  from datasets import load_dataset  
  ds = load_dataset("trojblue/million-song-subset")
  ```
- Explore and analyze various musical attributes easily.

## License
- **Original License**: Refer to the [Million Song Dataset license](http://millionsongdataset.com/pages/terms-of-use/)
- **Processed Version**: Shared for research and non-commercial purposes.

For more details, visit the [Million Song Dataset website](http://millionsongdataset.com).


## Appendix: Processing Code

The dataset was converted using the following snippet:

```python
import os
import unibox as ub
import pandas as pd
import numpy as np
import h5py
from tqdm import tqdm
from concurrent.futures import ProcessPoolExecutor

# https://github.com/tbertinmahieux/MSongsDB/blob/0c276e289606d5bd6f3991f713e7e9b1d4384e44/PythonSrc/hdf5_getters.py
import hdf5_getters

# Define dataset path
dataset_path = "/lv0/yada/dataproc5/data/MillionSongSubset"

# Function to extract all available fields from an HDF5 file
def extract_song_data(file_path):
    """Extracts all available fields from an HDF5 song file using hdf5_getters."""
    song_data = {}

    try:
        with hdf5_getters.open_h5_file_read(file_path) as h5:
            # Get all getter functions from hdf5_getters
            getters = [func for func in dir(hdf5_getters) if func.startswith("get_")]

            for getter in getters:
                try:
                    # Dynamically call each getter function
                    value = getattr(hdf5_getters, getter)(h5)

                    # Optimize conversions
                    if isinstance(value, np.ndarray):
                        value = value.tolist()
                    elif isinstance(value, bytes):
                        value = value.decode()

                    # Store in dictionary with a cleaned-up key name
                    song_data[getter[4:]] = value

                except Exception:
                    continue  # Skip errors but don't slow down

    except Exception as e:
        print(f"Error processing {file_path}: {e}")
    
    return song_data

# Function to process multiple files in parallel
def process_files_in_parallel(h5_files, num_workers=8):
    """Processes multiple .h5 files in parallel."""
    all_songs = []

    with ProcessPoolExecutor(max_workers=num_workers) as executor:
        for song_data in tqdm(executor.map(extract_song_data, h5_files), total=len(h5_files)):
            if song_data:
                all_songs.append(song_data)
    
    return all_songs

# Find all .h5 files
h5_files = [os.path.join(root, file) for root, _, files in os.walk(dataset_path) for file in files if file.endswith(".h5")]

# Process files in parallel
all_songs = process_files_in_parallel(h5_files, num_workers=24)

# Convert to Pandas DataFrame
df = pd.DataFrame(all_songs)

ub.saves(df, "hf://trojblue/million-song-subset", private=False)
```