Lakh MIDI Model

Model Overview

This repository contains the retrained model based on the original code and architecture provided by AI-Guru/MMM-JSB. The model has been trained from scratch on the Lakh MIDI Dataset, which has been carefully cleaned and prepared for this task.

The model is suitable for generating MIDI sequences and offers enhanced performance due to the improved dataset and careful retraining. The training process took approximately 50 hours on an RTX 4080 Super GPU, utilizing a dataset of about 6,000 MIDI compositions. The files included in this repository are essential for loading and utilizing the model efficiently.

Files in the Repository

The repository includes the following files:

config.json:
- Contains the configuration of the model architecture. This includes details such as the number of layers, hidden dimensions, attention heads, and other parameters used to define the model.
generation_config.json:
- Contains generation-specific settings, such as maximum sequence length, temperature, top-k, and top-p sampling parameters. These configurations are crucial for controlling the behavior of the MIDI sequence generation process.
model.safetensors:
- The model weights saved in the safetensors format for efficient and secure loading. This format ensures safe deserialization of model weights.
training_args.bin:
- Stores the training arguments and hyperparameters used during the training process. This file can be useful for reproducing the training setup or understanding the specifics of the training process.

Dataset Details

The model was trained on the Lakh MIDI Dataset, which has undergone extensive cleaning to ensure high-quality training data. The cleaning process involved:

Removing duplicates.
Ensuring proper formatting of MIDI files.
Filtering out noisy or incomplete data.

This dataset was chosen for its diverse range of MIDI sequences, providing the model with a rich set of training examples.

Original Code Base

The original model and architecture are based on the repository AI-Guru/MMM-JSB. This implementation has been retrained from scratch to work with the Lakh MIDI Dataset for MIDI generation tasks.

License

This model follows the licensing terms of the original repository. Please review the license for more details.