Code-LMs

Why we included this project

This is the kind of repository you bookmark when you want to run a code-generating model without fighting through research papers. It covers the full setup path for PolyCoder, from downloading checkpoints to loading them through HuggingFace or GPT-NeoX and generating code from prompts. The authors also document the exact evaluation procedure from their paper, including perplexity and HumanEval runs, which is useful if you are comparing models or reproducing results. There is practical detail on the multilingual training data covering 12 programming languages and on generation settings like temperature. It is not a polished tool but a working reference written by the people who trained the models, and that is exactly what makes it useful for experiments.