nintorac/ntrc_lmd_data

NTRC LMD Data

Check out the tables prefixed with mart_* to get started for some easy to understand and interesting data

This dataset contains an opionated, structured and lightly cleaned release of the Lakh MIDI Dataset by Colin Raffel, transformed using modern data engineering practices into a Data Vault 2.0 model.

Dataset Description

The original Lakh MIDI Dataset is a collection of 176,581 unique MIDI files, 45,129 of which have been matched and aligned to entries in the Million Song Dataset. This processed version transforms the raw data into a structured format suitable for music information retrieval research and analysis.

Key ideas in this release:

• Structured Data Vault 2.0 model with hubs, links, and satellites
• Bronze layer with raw extracted data in Parquet format (not part of this distribution, see https://github.com/Nintorac/ntrc_lmd to build it)
• Silver layer with cleaned, deduplicated, and structured data
• Gold layer with clean derormalized views ready for ML pipelines and other downstream usecases

Usage Examples

Connecting to the Remote Database

You can attach to the remote Lakh MIDI database using DuckDB and query the processed data directly:

-- Attach the remote database
ATTACH 'hf://datasets/nintorac/ntrc_lakh_midi/lakh_remote.duckdb' AS lakh_remote;

-- Sample the artist profile mart table
CREATE TABLE my_artist_sample AS 
SELECT * FROM lakh_remote.mart_artist_profile 
WHERE artist_profile_tier = 'High Profile' 
LIMIT 100;


-- Query the sample
SELECT 
    artist_name,
    total_tracks,
    avg_tempo,
    most_common_key,
    top_terms[1:3] as top_3_terms
FROM my_artist_sample
ORDER BY total_tracks DESC;

This approach allows you to fetch data from the gold layer, without having to manage the full silver layer locally. It's pretty slow though so it would be best to follow the above pattern to save a table locally before running analysis.

Data Structure

To understand the schema of the silver layer check out the schema_silver.dbml - load it in dbdiagram.io for an interactive view.

Below is a high level overview of how the tables fit together.

Silver Layer (silver_lakh_midi/)

Structured using Data Vault 2.0 methodology:

Hubs (Business Entities)

• hub_track.parquet - MusicBrainz tracks
• hub_artist.parquet - Artists (including similar artists)
• hub_release.parquet - 7digital releases
• hub_midi_file.parquet - MIDI files by MD5 hash
• hub_midi_source.parquet - Source file paths
• hub_key_signature.parquet - Musical key signatures (0-11)
• hub_mode.parquet - Musical modes (0=minor, 1=major)

Links (Relationships)

• link_track_midi.parquet - Track-to-MIDI matches
• link_track_artist.parquet - Track-to-artist relationships
• link_track_release.parquet - Track-to-release relationships
• link_artist_similar.parquet - Artist similarity relationships
• link_midi_source.parquet - MIDI-to-source path mappings

Satellites (Descriptive Data)

• sat_track/ - Track details, audio analysis, and time-series arrays (partitioned)
• sat_artist.parquet - Artist metadata and location info
• sat_release.parquet - Release information
• sat_midi_file/ - MIDI file content and size (partitioned)
• sat_match_scores.parquet - Match quality scores
• sat_artist_similarity.parquet - Similarity rankings
• sat_artist_terms.parquet - Echo Nest artist terms
• sat_artist_mbtags.parquet - MusicBrainz tags
• sat_key_signature.parquet - Human-readable key signature names
• sat_mode.parquet - Human-readable mode names

Data Quality Features

• Deduplication: Consistent handling of duplicate entities
• Referential Integrity: All foreign keys reference valid parents
• Partitioned Storage: Large tables partitioned by hash key prefix for scalable processing
• Comprehensive Testing: Data quality tests ensure consistency and completeness

Technical Details

• Format: Apache Parquet with snappy compression
• Schema: Data Vault 2.0 with hash keys for scalable joins
• Processing: Built with dlt, dbt, and DuckDB
• Partitioning: Large tables partitioned by first character of hash keys (0-9, a-f)

Original Dataset Information

This processed dataset is based on the Lakh MIDI Dataset, which contains:

• 176,581 unique MIDI files from LMD-full
• 45,129 matched files aligned to Million Song Dataset entries
• Audio analysis features from Echo Nest
• Artist metadata including MusicBrainz tags and similarity data

Citations & License

License

This dataset is distributed under the CC-BY 4.0 license, following the original Lakh MIDI Dataset license.

Required Citations

If you use this dataset, please cite the original Lakh MIDI Dataset:

@phdthesis{raffel2016learning,
  title={Learning-Based Methods for Comparing Sequences, with Applications to Audio-to-MIDI Alignment and Matching},
  author={Raffel, Colin},
  year={2016},
  school={Columbia University}
}

For the Million Song Dataset metadata, please also cite:

@inproceedings{bertin2011million,
  title={The Million Song Dataset},
  author={Bertin-Mahieux, Thierry and Ellis, Daniel PW and Whitman, Brian and Lamere, Paul},
  booktitle={Proceedings of the 12th International Society for Music Information Retrieval Conference},
  pages={591--596},
  year={2011}
}

Attribution Note

The original MIDI files were scraped from publicly-available sources on the internet and de-duplicated by Colin Raffel. While MIDI files have a built-in mechanism for attribution (the Copyright meta-event), it is not used consistently, so attributing each individual MIDI file to a particular author is not feasible.

Processing Pipeline

This dataset was processed using:

• dlt for data extraction from unstructured files
• dbt for data transformation and testing
• DuckDB as the analytics database engine
• Data Vault 2.0 methodology for scalable data modeling

For more information about the processing pipeline, see the project repository.

Resources

• Colin Raffel's project page.
• Million Song Dataset
• Echo Nest Schema