ronantakizawa/github-codereview

Code Review Dataset

A large-scale dataset of the best human-written code reviews from top GitHub repositories.

Each row captures a moment where a human code reviewer left an inline comment on a pull request, and the author subsequently modified the code in response.

The dataset also includes negative examples — code from the same PRs that passed review without comments — to help models learn when code is acceptable.

This provides a natural signal for training models to:

• —Generate code review comments given a code diff
• —Apply review feedback by modifying code based on reviewer suggestions
• —Understand code quality patterns across languages and projects
• —Know when not to comment — recognizing clean code that needs no changes

Key Features

• —167K+ positive triplets from 725 top GitHub repositories
• —51K+ negative examples (~23% of dataset) of clean code labeled "No issues found."
• —37 programming languages (Python, TypeScript, Go, Rust, C++, JavaScript, C#, Java, Kotlin, Swift, and more)
• —Human-only reviews: AI/bot reviewers (Copilot, linter bots, etc.) are excluded
• —Quality-filtered: noise and auto-generated content removed
• —Chunk-focused: ~50 lines of context around the reviewed code, not entire files
• —Permissive licenses only: all source repos use MIT, Apache-2.0, BSD, or similar licenses
• —Verified changes: only includes triplets where the code chunk actually changed after the review

Collection Methodology

1. 1.Repo selection: Top GitHub repos by stars with permissive licenses, sourced from ronantakizawa/github-top-projects and curated additions
2. 2.PR discovery: Paginate merged PRs, filter bot authors, fetch inline review comments
3. 3.Comment filtering: Remove bots, noise patterns, auto-generated comments, non-English text, non-code files, reply comments
4. 4.Triplet extraction: Fetch file contents at the review commit (before) and PR head (after), extract focused chunks around the comment line
5. 5.Change verification: Only keep triplets where the code chunk around the comment actually changed
6. 6.Negative extraction: For each reviewed PR, identify source code files that were changed but received no review comments; extract a ~50-line chunk as a negative example labeled "No issues found."

Splits

Splits are deterministic by repository — all examples from the same repo appear in the same split.

Schema

Usage

from datasets import load_dataset

ds = load_dataset("ronantakizawa/github-codereview")

# Get a training example
example = ds["train"][0]
print(f"Review comment: {example['reviewer_comment']}")
print(f"Language: {example['language']}")
print(f"Before:\n{example['before_code'][:200]}")
print(f"After:\n{example['after_code'][:200]}")

Filter by language

python_reviews = ds["train"].filter(lambda x: x["language"] == "Python")

Filter by quality

high_quality = ds["train"].filter(lambda x: x["quality_score"] >= 0.5)

Positive examples only

positives = ds["train"].filter(lambda x: not x["is_negative"])

Negative examples only

negatives = ds["train"].filter(lambda x: x["is_negative"])

Citation

If you use this dataset, please cite:

@dataset{takizawa2026codereviewdiffs,
  title={Code Review Diffs: A Large-Scale Dataset of Review-Driven Code Changes},
  author={Takizawa, Ronan},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/ronantakizawa/github-codereview}
}