Skip to content
Showcase/SWE-bench: Code Agent Solution Evaluation
advancedevaluation

SWE-bench: Code Agent Solution Evaluation

Evaluation of code agent solutions to real GitHub issues. Annotators review code patches generated by AI agents, assess correctness, check test compatibility, and evaluate code quality.

Submit

設定ファイルconfig.yaml

# SWE-bench: Code Agent Solution Evaluation
# Based on "SWE-bench: Can Language Models Resolve Real-World GitHub Issues?" (Jimenez et al., ICLR 2024)
# Task: Review code patches generated by AI agents for real GitHub issues

annotation_task_name: "SWE-bench Code Agent Evaluation"
task_dir: "."

# Data configuration
data_files:
  - sample-data.json
item_properties:
  id_key: "id"
  text_key: "text"

# Output
output_annotation_dir: "annotation_output/"
output_annotation_format: "json"

# Display layout showing issue, patch, and test results
html_layout: |
  <div class="swebench-container">
    <div class="repo-info" style="background: #e8eaf6; padding: 10px; border-radius: 8px; margin-bottom: 15px;">
      <strong>Repository:</strong> {{repo_name}}
    </div>
    <div class="issue-section" style="background: #e8f5e9; padding: 15px; border-radius: 8px; margin-bottom: 15px;">
      <h3 style="margin-top: 0;">GitHub Issue:</h3>
      <div class="issue-text" style="font-size: 15px; white-space: pre-wrap;">{{text}}</div>
    </div>
    <div class="patch-section" style="background: #f5f5f5; padding: 15px; border-radius: 8px; margin-bottom: 15px; border: 2px solid #424242;">
      <h3 style="margin-top: 0;">Agent-Generated Patch:</h3>
      <pre style="white-space: pre-wrap; font-family: monospace; font-size: 13px; line-height: 1.5; overflow-x: auto;">{{patch}}</pre>
    </div>
    <div class="test-section" style="background: #fff3e0; padding: 15px; border-radius: 8px; border: 2px solid #f9a825;">
      <h3 style="margin-top: 0; color: #f9a825;">Test Results:</h3>
      <pre style="white-space: pre-wrap; font-family: monospace; font-size: 13px;">{{test_results}}</pre>
    </div>
  </div>

# Annotation schemes
annotation_schemes:
  # Correctness assessment
  - name: "correctness"
    description: "Does the patch correctly fix the reported issue?"
    annotation_type: radio
    labels:
      - "Correct - fully fixes the issue"
      - "Partially Correct - fixes some aspects but not all"
      - "Incorrect - does not fix the issue"
      - "Introduces new bugs - fix creates other problems"
    keyboard_shortcuts:
      "Correct - fully fixes the issue": "1"
      "Partially Correct - fixes some aspects but not all": "2"
      "Incorrect - does not fix the issue": "3"
      "Introduces new bugs - fix creates other problems": "4"

  # Code quality assessment
  - name: "code_quality"
    description: "Rate the overall quality of the generated code."
    annotation_type: radio
    labels:
      - "Clean - well-structured, follows conventions, maintainable"
      - "Acceptable - works but could be improved"
      - "Poor - messy, violates conventions, hard to maintain"
    keyboard_shortcuts:
      "Clean - well-structured, follows conventions, maintainable": "a"
      "Acceptable - works but could be improved": "s"
      "Poor - messy, violates conventions, hard to maintain": "d"

  # Review comments
  - name: "review_comments"
    description: "Provide code review comments explaining your assessment. Note specific issues, suggestions for improvement, or edge cases."
    annotation_type: text
    min_length: 20
    max_length: 600
    placeholder: "Provide specific code review feedback: correctness issues, style concerns, edge cases missed, test coverage gaps..."

  # Test compatibility
  - name: "test_compatibility"
    description: "Are the existing tests passing with this patch?"
    annotation_type: radio
    labels:
      - "All tests pass"
      - "Most tests pass (minor failures)"
      - "Some tests fail"
      - "Many tests fail"
      - "Cannot determine from provided info"
    keyboard_shortcuts:
      "All tests pass": "z"
      "Most tests pass (minor failures)": "x"
      "Some tests fail": "c"
      "Many tests fail": "v"
      "Cannot determine from provided info": "b"

# User configuration
allow_all_users: true

# Task assignment
instances_per_annotator: 50
annotation_per_instance: 2

サンプルデータsample-data.json

[
  {
    "id": "swe_001",
    "text": "Issue #4521: `DataFrame.groupby().apply()` raises TypeError when the applied function returns a scalar value for groups with a single row.\n\nSteps to reproduce:\n```python\nimport pandas as pd\ndf = pd.DataFrame({'key': ['a', 'b', 'b'], 'value': [1, 2, 3]})\nresult = df.groupby('key').apply(lambda x: x['value'].sum())\n```\nExpected: Returns a Series with group sums.\nActual: TypeError: 'int' object is not iterable",
    "patch": "--- a/pandas/core/groupby/groupby.py\n+++ b/pandas/core/groupby/groupby.py\n@@ -1245,7 +1245,10 @@ class GroupBy:\n     def _apply_func(self, func, data):\n         result = func(data)\n-        if isinstance(result, DataFrame):\n+        if is_scalar(result):\n+            return Series([result], index=data.index[:1])\n+        elif isinstance(result, DataFrame):\n             return result\n         elif isinstance(result, Series):\n             return result\n         else:\n             return Series(result)",
    "repo_name": "pandas-dev/pandas",
    "test_results": "PASSED: test_groupby_apply_scalar_return (3 subtests)\nPASSED: test_groupby_apply_dataframe_return\nPASSED: test_groupby_apply_series_return\nFAILED: test_groupby_apply_empty_groups - AssertionError: Expected empty DataFrame, got Series\n\nTotal: 15 passed, 1 failed, 0 errors"
  },
  {
    "id": "swe_002",
    "text": "Issue #8734: `json.dumps()` does not properly handle `datetime.date` objects, raising `TypeError: Object of type date is not JSON serializable`.\n\nThis should be handled gracefully with a default serializer or a clear error message suggesting a solution.",
    "patch": "--- a/Lib/json/encoder.py\n+++ b/Lib/json/encoder.py\n@@ -178,6 +178,8 @@ class JSONEncoder:\n     def default(self, o):\n         if isinstance(o, datetime.datetime):\n             return o.isoformat()\n+        elif isinstance(o, datetime.date):\n+            return o.isoformat()\n         raise TypeError(f\"Object of type {type(o).__name__} is not JSON serializable. \"\n                        f\"Consider using a custom encoder or converting to a \"\n                        f\"serializable type first.\")",
    "repo_name": "python/cpython",
    "test_results": "PASSED: test_date_serialization\nPASSED: test_datetime_serialization\nPASSED: test_custom_encoder_still_works\nPASSED: test_error_message_improved\nPASSED: test_nested_date_objects\n\nTotal: 12 passed, 0 failed, 0 errors"
  }
]

// ... and 8 more items

このデザインを取得

View on GitHub

Clone or download from the repository

クイックスタート：

git clone https://github.com/davidjurgens/potato-showcase.git
cd potato-showcase/agentic/swebench-code-agent-eval
potato start config.yaml

詳細

アノテーションタイプ

radiotext

ドメイン

Software EngineeringCode AgentsEvaluation

ユースケース

Code ReviewAgent EvaluationPatch Assessment

タグ

swe-benchcode-agentgithub-issuespatch-reviewsoftware-engineeringevaluation

問題を見つけた場合やデザインを改善したい場合は？

Issueを作成

関連デザイン

Code Review Annotation (CodeReviewer)

Annotation of code review activities based on the CodeReviewer benchmark. Annotators identify issues in code diffs, classify defect types, assign severity levels, make review decisions, and provide natural language review comments, supporting research in automated code review and software engineering.

spanradio

FActScore: Fine-grained Atomic Evaluation of Factual Precision

Atomic fact evaluation in LLM-generated text. Annotators decompose generated text into atomic facts and verify each fact as supported, not-supported, or irrelevant against a reference source. Based on the FActScore framework for evaluating factual precision in long-form text generation.

radiotext

GPQA - Graduate-Level Expert QA Evaluation

Expert-level question answering evaluation on graduate-level science questions from the GPQA benchmark (Rein et al., ICLR 2024). Questions span physics, chemistry, and biology, designed to be answerable only by domain experts.

numberradio