Automating Code Reviews Using Large Language Models
3 points by aravindputrevu 7 hours ago | 1 comments
  • aravindputrevu 7 hours ago |
    Stumbled on this piece while searching for Code Review on Google.

    > Microsoft Research and LinkedIn researchers have open-sourced CodeReviewer, a pre-trained transformer model that can automatically assess code changes, generate review comments, and suggest fixes. Trained on 7.9M pull requests across 9 programming languages, it achieves a 71.5% F1 score in identifying problematic code changes and can generate relevant review comments with 3.6/5.0 informativeness rating from human evaluators.

    > Unlike existing code models, CodeReviewer is specifically trained on code diffs and real-world review comments from high-quality GitHub repositories. The model outperforms previous approaches by learning to "think" like a code reviewer rather than just understanding source code.

    Technical details and model available at: https://github.com/microsoft/CodeBERT/tree/master/CodeReview...

    I'm thinking GitHub is gonna use some of this learning in Copilot?