GPT Training Tool

This repository contains a GPT Training Tool adapted from Andrej Karpathy's MiniGPT repository. The tool runs as an API and can build, train, test, store, and reload various GPT models. It will also generate a graph of your model architecture and provide a summary of its size. Note: Currently it is equipped to learn character encodings, but it can be easily updated to learn subword encodings.

Installation

Before using the GPT Training Tool, make sure to install the required packages:

pip install -r requirements.txt

Usage

To run the GPT Training Tool, follow these steps:

Start the Application

Run the main script:

python app.py

Endpoints

The application exposes the following endpoints:

• \new_model: creates a new model using provided params
• \load_model: loads the provided model from checkpoint
• \save_model: saves the current model weights, hyperparameters, and logs
• \get_model: gets the name of the currently loaded model
• \view_model: shows a summary of the model size
• \get_params: gets the list of the currently loaded params
• \update_params: patches the params file, currently only accepts training params
• \train: trains the currently loaded model
• \evaluate: evaluates the current loss on the test data set
• \generate: returns generation
• \complete: returns completion from prompt

A postman collection with all of these endpoints set up is stored in configs.