Tokenized MIDI Music Training Shards for LMD

Shards of the Lakh MIDI Dataset (LMD) processed into a tokenized format for machine learning training. The dataset was prepared by author 'choihy' and was last updated on Hugging Face in April 2026. It contains grouped-sample data packed into shards for efficient distribution and loading.

Use Cases

• Training autoregressive models on tokenized MIDI sequences to generate new musical compositions.
• Fine-tuning language models on the token and pitch/rhythm (PR) features for music style transfer tasks.
• Benchmarking sequence modeling performance on structured, tokenized music data from the LMD corpus.
• Preprocessing pipeline development using the provided unpacking script to restore the raw grouped-sample directory tree.

Strengths

• Data is structured into shards for efficient distributed training workflows.
• Includes a dedicated unpacking script (lmd_training_shards_unpack.py) to restore the expected directory format.

Limitations

• The specific number of shards, total rows of data, and file size are not provided.
• The raw description is incomplete, requiring a visit to the external dataset page for full details on contents.

Provenance

Source

Processed from the Lakh MIDI Dataset (LMD).

Collection Method

MIDI files tokenized and packed into grouped-sample shards.

Freshness

Last updated on the platform in April 2026.