metadata

license:
  - apache-2.0
size_categories:
  - 100<n<1K
pretty_name: EditBench
tags:
  - code
  - code-editing
  - code-generation
metrics:
  - execution-accuracy

EditBench Dataset

This dataset contains code editing tasks extracted from the EditBench evaluation framework specifically designed for evaluating model performance on code editing tasks. It is provided as a test-only benchmark. Each sample includes:

Core Files (Python)

original_code.py: Starting code file
highlighted_code.py: Specific section of code to be modified
instruction.txt: User instructions for the task
test_code.py: Tests that validate the implementation

Supporting Files (Python)

requirements.txt: Dependencies needed to run the code
conftest.py: Pytest configuration
test_utils.py: Utilities for testing

Core Files (JavaScript)

original_code.js: Starting code file (or .jsx)
highlighted_code.js: Specific section of code to be modified
instruction.txt: User instructions for the task
test_code: Tests that validate the implementation (from tests/*.test.js)
package_json: NPM package configuration
jest_setup: Jest testing setup (if applicable)
babel_config: Babel configuration (if applicable)
other_files: Additional files needed for the project

Dataset Statistics

Total samples: 113
Python samples: 104
JavaScript samples: 9
Expected samples: 113 (57 easy + 56 hard questions)
Found samples: 113 / 113

Usage

This dataset is provided as a test-only benchmark and can be loaded directly with the Hugging Face Datasets library:

from datasets import load_dataset

# Note that this dataset only has a 'test' split
dataset = load_dataset("your-username/editbench", split="test")

Ethical Considerations and Limitations

This dataset is provided exclusively for benchmark/evaluation purposes
Models should NOT be trained on this dataset, as it is specifically designed to test model capabilities
Hugging Face's Terms of Service prohibit using benchmark datasets for training
We recommend implementing your model's training pipeline to explicitly exclude this dataset

Citation

If you use this dataset, please cite the original EditBench work.

@misc{chi2025editbench,
  title        = {EditBench: Evaluating LLM Abilities to Perform Real-World Code Edits},
  author       = {Wayne Chi and Valerie Chen and Ryan Shar and Aditya Mittal and Jenny Liang and Wei-Lin Chiang and Anastasios Nikolas Angelopoulos and Ion Stoica and Graham Neubig and Ameet Talwalkar and Chris Donahue},
  year         = {2025},
  note         = {arXiv preprint}
}

Usage Restrictions

This dataset is provided for research and evaluation purposes only. By using this dataset, you agree not to:

Train models on it (it is a benchmark dataset)
Scrape or incorporate it into pretraining data
Use it for any purpose other than evaluation