GPyT Project

GPyT is a GPT2 model trained from scratch (not fine tuned) on Python code from Github. Overall, it was ~200GB of pure Python code, the current GPyT model is a mere 2 epochs through this data, so it may benefit greatly from continued training and/or fine-tuning.

Newlines are replaced by

Input to the model is code, up to the context length of 1024, with newlines replaced by

Here's a quick example of using this model:

from transformers import AutoTokenizer, AutoModelWithLMHead

tokenizer = AutoTokenizer.from_pretrained("Reverb/GPyT")
model = AutoModelWithLMHead.from_pretrained("Reverb/GPyT")

# copy and paste some code in here
inp = """import"""

newlinechar = "<N>"
converted = inp.replace("\n", newlinechar)
tokenized = tokenizer.encode(converted, return_tensors='pt')
resp = model.generate(tokenized)

decoded = tokenizer.decode(resp[0])
reformatted = decoded.replace("<N>","\n")

print(reformatted)

Should produce:

import numpy as np
import pytest

import pandas as pd<N

The Journey

The model took 6 major steps which are:

Data Collection
Raw Data Cleaning
Data Preprocessing
Building & Training the Tokenizer
Testing the Model on Large Dataset
Deploying the Final Model on HuggingFace

Data Collection

The data was collected from python github repositories using web scraping techniques, It took nearly a day to gather 200GB worth of data.

Raw Data Cleaning

200GB of python code?? sounds ridiculous! that's why we needed to clean the downloaded repositories from any non-python files such as PDF,idx..etc

Data Preprocessing

I tried splitting the lines of code for each repository then merged them all under one single text file named python_text_data.txt

Building & Training the Tokenizer

For this step I have used ByteLevelBPETokenizer and trained it then saved the model on the desktop

Testing the Model on Large Dataset

After training the tokenizer on a large dataset, It was time for some tests to see how good is the model before proceeding.

Considerations:

This model is intended for educational and research use only. Do not trust model outputs.

Model is highly likely to regurgitate code almost exactly as it saw it. It's up to you to determine licensing if you intend to actually use the generated code.

All Python code was blindly pulled from github. This means included code is both Python 2 and 3, among other more subtle differences, such as tabs being 2 spaces in some cases and 4 in others...and more non-homologous things.

Along with the above, this means the code generated could wind up doing or suggesting just about anything. Run the generated code at own risk...it could be anything