differentiablepizza / twitch-simulator

GPT-Neo fine tuning from Twitch chat conversations

Introduction

These are scripts for treatment, training, evaluation of a fine tuned version of GPT-Neo from data of Twitch Chat logs and streamer transcribed conversations. Currently, the successfully trained models are being saved on HuggingFace.

Instalation

This project is known to work with Python 3.10+. To install all necessary packages, open a terminal and run the following command in the root directory:

pip install -r requirements.txt

Depending on your PIP installation, you'll need to use pip3 isntead of pip. This usually occurs when you have more than one version of Python installed on your computer (2.7 and 3.10 for example).

Training

The model can be trained with the file model-neo.py, the way it is currently set up requires a GPU with Cuda enabled. In case you don't have a capable GPU or the model ran out of memory, you can also try to run the model directly on your CPU at the cost of a longer training time. For that, comment out lines with .to('cuda') and set no_cuda to True and comment out the parameter fp16 on the TrainingArguments constructor.

Then, execute the trainer as a notebook or do

python model-neo.py

Evaluation

The file evaluator.py is able to generate text based on the input model. Make sure that model_name is set to the correct path and model version that you want to use. This script doesn't require a GPU to function. The input prompt is the text contained prompt, which you can change at will. The response is printed out at the end.

For evaluation, execute the file as a notebook or do

python evaluator.py