Similar to GitHub Copilot with one major difference: Code search is leveraged to make up for smaller models and less data. The generation model makes use of search results along with code context to generate contextual code. Moreover, code search results from all over GitHub are displayed in a seperate file along with the exact url of the source code.
Currently, GPT-Neo-125M model is used for generation.
Training code is under ‘gen’ directory.
Code to finetune the model and setup ScaNN is under ‘search’ directory.
(Note the dot ‘.’ at the end.)
# concat two dicts.
// merge two arrays.
VSCode 1.59.0 or up.
If you spot any mistakes or any possible improvements, please feel free to let me know and contributions are welcome!