metadata
language: Python
tags:
- Code
- GPyT
- code generator
license: MIT
GPyT is a GPT2 model trained from scratch (not fine tuned) on Python code from Github. Overall, it was ~80GB 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 <N>
Input to the model is code, up to the context length of 1024, with newlines replaced by <N>
Here's an example of a quick converter to take your multi-line code and replace the newlines:
inp = """def do_something():
print("Hello")
"""
newlinechar = "<N>"
converted = inp.replace("\n", newlinechar)
print("length:", len(converted))
print(converted)
This should give you something like:
def do_something():<N> print("Hello")<N>
...which is what the model is expecting as input.
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