Datasets:

Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
Size:
< 1K
ArXiv:
Tags:
code
Libraries:
Datasets
Dask
License:
Dataset Viewer
The dataset viewer is not available for this dataset.
Unexpected token '<', "<html> <h"... is not valid JSON

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Dataset Card for EffiBench-X

EffiBench-X is the first multi-language benchmark designed specifically to evaluate the efficiency of LLM-generated code across six programming languages: Python, C++, Java, JavaScript, Ruby, and Golang. The dataset comprises 623 competitive programming problems paired with human expert solutions as efficiency baselines.

Dataset Details

Dataset Description

EffiBench-X addresses critical limitations in existing code generation benchmarks by providing:

  • Multi-language evaluation across Python, C++, Java, JavaScript, Ruby, and Golang

  • Efficiency-focused metrics including execution time, memory peak, and memory integral

  • Human expert baselines for reliable efficiency comparison

  • Curated by: Yuhao Qing, Boyu Zhu, Mingzhe Du, Zhijiang Guo, Terry Yue Zhuo, Qianru Zhang, Jie M. Zhang, Heming Cui, Siu-Ming Yiu, Dong Huang, See-Kiong Ng, Luu Anh Tuan

  • Institutions: HKU, UCL, NTU, NUS, HKUST, Monash University, CSIRO's Data61, KCL

  • Language(s) (NLP): English

  • License: Apache License 2.0

Dataset Sources

Uses

Direct Use

  • Benchmarking LLM code generation efficiency: Evaluate models on runtime performance, memory usage, and correctness across multiple languages
  • Cross-language performance analysis: Compare model capabilities across different programming paradigms
  • Model development: Train and fine-tune models for efficient code generation
  • Research: Study efficiency gaps between LLM-generated and human expert code

Out-of-Scope Use

  • Production deployment without validation: Solutions should be verified before production use
  • Security-critical applications: The dataset focuses on algorithmic efficiency, not security
  • Non-competitive programming domains: Problems are algorithmic in nature and may not represent all software engineering contexts

Dataset Structure

The dataset contains 623 problems categorized into:

  • Functional problems: Implement specific functions/classes with I/O handled by test templates
  • Standard I/O problems: Complete programs reading from stdin and writing to stdout

Key fields per record include:

  • id, title, title_slug, description, description_md, difficulty, tags, source, url, type
  • Limits: time_limit_nanos, memory_limit_bytes
  • Code artifacts:
    • starter_code: language-keyed starter snippets
    • solutions: language-keyed canonical solutions (e.g., for cpp, golang, java, javascript, python3, ruby)
    • test_case_generator: executable code string that programmatically produces tests
    • evaluator: executable code string to determine pass/fail given expected vs. program output
    • generated_tests: serialized tests produced by the generator
    • test_runners: language-keyed runner templates for executing solutions

All problems are from competitive programming platforms.

Dataset Creation

Curation Rationale

Existing code generation benchmarks primarily focus on functional correctness with limited attention to efficiency, often restricted to Python. EffiBench-X addresses three critical limitations:

  1. Language diversity: Extends beyond Python to include statically-typed (C++, Java, Go) and dynamically-typed languages (Python, JavaScript, Ruby)
  2. Data contamination: Uses recent problems (post-October 2023) to avoid memorization effects
  3. Complexity: Features algorithmically challenging problems requiring optimization techniques

Source Data

Data Collection and Processing

Problems are curated from competitive programming platforms. Each problem includes:

  • Human expert solutions verified for correctness and efficiency
  • 100 programmatically generated test cases
  • Test runners and evaluators for automated assessment
  • Cross-language validation to ensure consistency

Who are the source data producers?

  • Problem creators: Competitive programming platforms and contest organizers
  • Solution authors: Human expert programmers from competitive programming communities
  • Dataset curators: EffiBench research team

Citation

Please cite our paper if you use this dataset:

@article{qing2025effibench,
  title={EffiBench-X: A Multi-Language Benchmark for Measuring Efficiency of LLM-Generated Code},
  author={Qing, Yuhao and Zhu, Boyu and Du, Mingzhe and Guo, Zhijiang and Zhuo, Terry Yue and Zhang, Qianru and Zhang, Jie M and Cui, Heming and Yiu, Siu-Ming and Huang, Dong and Ng, See-Kiong and Tuan, Luu Anh},
  journal={Advances in neural information processing systems},
  year={2025}
}

More Information

  • Dataset Statistics: 623 problems, 100 test cases per problem, 6 languages
  • Evaluation: Sandboxed execution environment for consistent performance measurements
  • For detailed information and benchmark results, please refer to the paper and GitHub repository

Dataset Card Contact

For questions and feedback, please open an issue on our GitHub repository.

Downloads last month
129

Paper for EffiBench/effibench-x