ronantakizawa/github-codereview

---
license: other
task_categories:
- text-generation
language:
- en
- code
tags:
- code-review
- code-generation
- software-engineering
- pull-requests
- github
size_categories:
- 100K<n<1M
---

# 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. **Repo selection**: Top GitHub repos by stars with permissive licenses, sourced from [ronantakizawa/github-top-projects](https://huggingface.co/datasets/ronantakizawa/github-top-projects) and curated additions
2. **PR discovery**: Paginate merged PRs, filter bot authors, fetch inline review comments
3. **Comment filtering**: Remove bots, noise patterns, auto-generated comments, non-English text, non-code files, reply comments
4. **Triplet extraction**: Fetch file contents at the review commit (before) and PR head (after), extract focused chunks around the comment line
5. **Change verification**: Only keep triplets where the code chunk around the comment actually changed
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

| Split | Percentage | Description |
|-------|-----------|-------------|
| train | 90% | Training data |
| test | 5% | Test data |
| validation | 5% | Validation data |

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

## Schema

| Column | Type | Description |
|--------|------|-------------|
| `pr_title` | string | Pull request title |
| `pr_number` | int | PR number |
| `repo_name` | string | Full repo name (owner/repo) |
| `repo_stars` | int | GitHub stars |
| `repo_language` | string | Primary repo language |
| `author_username` | string | PR author's GitHub username |
| `reviewer_username` | string | Reviewer's GitHub username |
| `before_code` | string | ~50 lines of code around the comment, before the fix |
| `reviewer_comment` | string | The inline review comment text (or "No issues found." for negatives) |
| `after_code` | string | ~50 lines of code around the comment, after the fix |
| `diff_context` | string | The PR diff hunk where the comment was placed |
| `file_path` | string | File path within the repo |
| `comment_line` | int | Line number within the code chunk (0 for negatives) |
| `language` | string | Programming language |
| `quality_score` | float | Comment quality score (0.0-1.0; 1.0 for negatives) |
| `comment_type` | string | Category: suggestion, question, nitpick, bug, refactor, style, security, performance, none |
| `comment_length` | int | Character count of reviewer comment |
| `before_lines` | int | Line count of before code |
| `after_lines` | int | Line count of after code |
| `is_negative` | bool | True if this is a negative example (no reviewer comment) |

## Usage

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

### Filter by quality

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

### Positive examples only

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

### Negative examples only

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

## Citation

If you use this dataset, please cite:

```bibtex
@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}
}
```