SWE-CARE: A Comprehensiveness-aware Benchmark for Code Review Evaluation

Dataset Description

SWE-CARE (Software Engineering - Comprehensive Analysis and Review Evaluation) is a comprehensiveness-aware benchmark for evaluating Large Language Models (LLMs) on repository-level code review tasks. The dataset features real-world code review scenarios from popular open-source Python and Java repositories, with comprehensive metadata and reference review comments.

Dataset Summary

• Repository: inclusionAI/SWE-CARE
• Paper: CodeFuse-CR-Bench: A Comprehensiveness-aware Benchmark for End-to-End Code Review Evaluation
• Languages: Python
• License: Apache 2.0
• Splits:
  • test: 671 instances (primary evaluation set)
  • dev: 7,086 instances (development/training set)

Dataset Structure

Data Instances

Each instance in the dataset represents a code review task with the following structure:

{
    "instance_id": "voxel51__fiftyone-2353@02e9ba1",
    "repo": "voxel51/fiftyone",
    "language": "Python",
    "pull_number": 2353,
    "title": "Fix issue with dataset loading",
    "body": "This PR fixes...",
    "created_at": "2023-01-15T10:30:00Z",
    "problem_statement": "Issue #2350: Dataset fails to load...",
    "hints_text": "Comments from the issue discussion...",
    "resolved_issues": [
        {
            "number": 2350,
            "title": "Dataset loading error",
            "body": "When loading datasets..."
        }
    ],
    "base_commit": "abc123...",
    "commit_to_review": {
        "head_commit": "def456...",
        "head_commit_message": "Fix dataset loading logic",
        "patch_to_review": "diff --git a/file.py..."
    },
    "reference_review_comments": [
        {
            "text": "Consider adding error handling here",
            "path": "src/dataset.py",
            "diff_hunk": "@@ -10,5 +10,7 @@...",
            "line": 15,
            "start_line": 14,
            "original_line": 15,
            "original_start_line": 14
        }
    ],
    "merged_commit": "ghi789...",
    "merged_patch": "diff --git a/file.py...",
    "metadata": {
        "problem_domain": "Bug Fixes",
        "difficulty": "medium",
        "estimated_review_effort": 3
    }
}

Data Fields

Core Fields

• instance_id (string): Unique identifier in format repo_owner__repo_name-PR_number@commit_sha_short
• repo (string): GitHub repository in format owner/name
• language (string): Primary programming language (Python or Java)
• pull_number (int): GitHub pull request number
• title (string): Pull request title
• body (string): Pull request description
• created_at (string): ISO 8601 timestamp of PR creation

Problem Context

• problem_statement (string): Combined title and body of resolved issue(s)
• hints_text (string): Relevant comments from issues prior to the PR
• resolved_issues (list): Array of resolved issues with:
  • number (int): Issue number
  • title (string): Issue title
  • body (string): Issue description

Code Changes

• base_commit (string): Base commit SHA before changes
• commit_to_review (dict): The commit being reviewed:
  • head_commit (string): Commit SHA to review
  • head_commit_message (string): Commit message
  • patch_to_review (string): Git diff of changes to review
• merged_commit (string): Final merged commit SHA
• merged_patch (string): Final merged changes (ground truth)

Reference Reviews

• reference_review_comments (list): Human code review comments with:
  • text (string): Review comment text
  • path (string): File path being reviewed
  • diff_hunk (string): Relevant code diff context
  • line (int): Line number in new version
  • start_line (int): Start line for multi-line comments
  • original_line (int): Line number in original version
  • original_start_line (int): Original start line

Metadata

• metadata (dict): LLM-classified attributes:
  • problem_domain (string): Category like "Bug Fix", "Feature", "Refactoring", etc.
  • difficulty (string): "Easy", "Medium", or "Hard"
  • estimated_review_effort (int): Scale of 1-5 for review complexity

Data Splits

Split	Instances	Description
test	671	Primary evaluation set for benchmarking
dev	7,086	Development set for training/fine-tuning

Usage

Loading the Dataset

from datasets import load_dataset

# Load the test split (default for evaluation)
dataset = load_dataset("inclusionAI/SWE-CARE", split="test")

# Load the dev split
dev_dataset = load_dataset("inclusionAI/SWE-CARE", split="dev")

# Load both splits
full_dataset = load_dataset("inclusionAI/SWE-CARE")

Using with SWE-CARE Evaluation Framework

from swe_care.utils.load import load_code_review_dataset

# Load from Hugging Face (default)
instances = load_code_review_dataset()

# Access instance data
for instance in instances:
    print(f"Instance: {instance.instance_id}")
    print(f"Repository: {instance.repo}")
    print(f"Problem: {instance.problem_statement}")
    print(f"Patch to review: {instance.commit_to_review.patch_to_review}")
    print(f"Reference comments: {len(instance.reference_review_comments)}")

Running Evaluation

See the GitHub repository for detailed documentation and examples.

Evaluation Metrics and Baselines Results

See the paper for comprehensive evaluation metrics and baseline results on various LLMs.

Additional Information

Citation

If you use this dataset in your research, please cite:

@misc{guo2025codefusecrbenchcomprehensivenessawarebenchmarkendtoend,
      title={CodeFuse-CR-Bench: A Comprehensiveness-aware Benchmark for End-to-End Code Review Evaluation in Python Projects}, 
      author={Hanyang Guo and Xunjin Zheng and Zihan Liao and Hang Yu and Peng DI and Ziyin Zhang and Hong-Ning Dai},
      year={2025},
      eprint={2509.14856},
      archivePrefix={arXiv},
      primaryClass={cs.SE},
      url={https://arxiv.org/abs/2509.14856}, 
}

Contributions

We welcome contributions! Please see our GitHub repository for:

• Data collection improvements
• New evaluation metrics
• Baseline model results
• Bug reports and feature requests

License

This dataset is released under the Apache 2.0 License. See LICENSE for details.

Changelog

• v0.2.0 (2025-10): Expanded dataset to 671 test instances
• v0.1.0 (2025-09): Initial release with 601 test instances and 7,086 dev instances