metadata
license: apache-2.0
language:
- code
tags:
- code-review
- codet5
- code-language-model
pipeline_tag: text2text-generation
codereviewer-msg-pyjs
microsoft/codereviewer fine-tuned on the CodeReviewer comment-generation
task (Python + JavaScript). Given a code change (a diff encoded with
<add>/<del>/<keep> line markers), it generates a natural-language review comment.
- Base model:
microsoft/codereviewer(CodeT5 family, pre-trained on diffs+reviews) - Task: review-comment generation (the CodeReviewer
msgtask) - Eval: official CodeReviewer smoothed BLEU-4 (
rmstop=True) = 7.26 on the Python/JS test split (reference paper ~5.73).
Input format
Prefix each diff line with an edit marker, e.g.:
<keep> def get_user(conn, username):
<del> cur.execute("SELECT * FROM users WHERE name = '" + username + "'")
<add> cur.execute("SELECT * FROM users WHERE name = ?", (username,))