MxEval

Multilingual Execution Evaluation

Dataset Summary

This repository contains data and code to perform execution-based multi-lingual evaluation of code generation capabilities and the corresponding data, namely, a multi-lingual benchmark MBXP, multi-lingual MathQA and multi-lingual HumanEval.
Results and findings can be found in the paper "Multi-lingual Evaluation of Code Generation Models".

Supported Tasks and Leaderboards

• MBXP
• Multi-HumanEval
• MathQA-X

Languages

The programming problems are written in multiple programming languages and contain English natural text in comments and docstrings.

Dataset Structure

To lookup currently supported datasets

get_dataset_config_names("mxeval/mxeval")
['mathqa-x', 'mbxp', 'multi-humaneval']

To load a specific dataset and language

from datasets import load_dataset
load_dataset("mxeval/mxeval", "mbxp", split="python")
Dataset({
    features: ['task_id', 'language', 'prompt', 'test', 'entry_point', 'description', 'canonical_solution'],
    num_rows: 974
})

Data Instances

An example of a dataset instance:

{
  "task_id": "MBSCP/6",
  "language": "scala",
  "prompt": "object Main extends App {\n    /**\n     * You are an expert Scala programmer, and here is your task.\n     * * Write a Scala function to check whether the two numbers differ at one bit position only or not.\n     *\n     * >>> differAtOneBitPos(13, 9)\n     * true\n     * >>> differAtOneBitPos(15, 8)\n     * false\n     * >>> differAtOneBitPos(2, 4)\n     * false\n     */\n    def differAtOneBitPos(a : Int, b : Int) : Boolean = {\n",
  "test": "\n\n    var arg00 : Int = 13\n    var arg01 : Int = 9\n    var x0 : Boolean = differAtOneBitPos(arg00, arg01)\n    var v0 : Boolean = true\n    assert(x0 == v0, \"Exception -- test case 0 did not pass. x0 = \" + x0)\n\n    var arg10 : Int = 15\n    var arg11 : Int = 8\n    var x1 : Boolean = differAtOneBitPos(arg10, arg11)\n    var v1 : Boolean = false\n    assert(x1 == v1, \"Exception -- test case 1 did not pass. x1 = \" + x1)\n\n    var arg20 : Int = 2\n    var arg21 : Int = 4\n    var x2 : Boolean = differAtOneBitPos(arg20, arg21)\n    var v2 : Boolean = false\n    assert(x2 == v2, \"Exception -- test case 2 did not pass. x2 = \" + x2)\n\n\n}\n",
  "entry_point": "differAtOneBitPos",
  "description": "Write a Scala function to check whether the two numbers differ at one bit position only or not."
}

Data Fields

• task_id: identifier for the data sample
• prompt: input for the model containing function header and docstrings
• canonical_solution: solution for the problem in the prompt
• description: task description
• test: contains function to test generated code for correctness
• entry_point: entry point for test
• language: programming lanuage identifier to call the appropriate subprocess call for program execution

Data Splits

• HumanXEval
  • Python
  • Java
  • JavaScript
  • Csharp
  • CPP
  • Go
  • Kotlin
  • PHP
  • Perl
  • Ruby
  • Swift
  • Scala
• MBXP
  • Python
  • Java
  • JavaScript
  • TypeScript
  • Csharp
  • CPP
  • Go
  • Kotlin
  • PHP
  • Perl
  • Ruby
  • Swift
  • Scala
• MathQA
  • Python
  • Java
  • JavaScript

Dataset Creation

Curation Rationale

Since code generation models are often trained on dumps of GitHub a dataset not included in the dump was necessary to properly evaluate the model. However, since this dataset was published on GitHub it is likely to be included in future dumps.

Personal and Sensitive Information

None.

Social Impact of Dataset

With this dataset code generating models can be better evaluated which leads to fewer issues introduced when using such models.

Dataset Curators

AWS AI Labs

Execution

Execution Example

Install the repo mbxp-exec-eval to execute generations or canonical solutions for the prompts from this dataset.

>>> from datasets import load_dataset
>>> from mxeval.execution import check_correctness
>>> mbxp_python = load_dataset("mxeval/mxeval", "mbxp", split="python")
>>> example_problem = mbxp_python[0]
>>> check_correctness(example_problem, example_problem["canonical_solution"], timeout=20.0)
{'task_id': 'MBPP/1', 'passed': True, 'result': 'passed', 'completion_id': None, 'time_elapsed': 10.582208633422852}

Considerations for Using the Data

Make sure to sandbox the execution environment since generated code samples can be harmful.

Licensing Information

LICENSE
THIRD PARTY LICENSES

Citation Information

@article{mbxp_athiwaratkun2022,
  title = {Multi-lingual Evaluation of Code Generation Models},
  author = {Athiwaratkun, Ben and
   Gouda, Sanjay Krishna and
   Wang, Zijian and
   Li, Xiaopeng and
   Tian, Yuchen and
   Tan, Ming
   and Ahmad, Wasi Uddin and
   Wang, Shiqi and
   Sun, Qing and
   Shang, Mingyue and
   Gonugondla, Sujan Kumar and
   Ding, Hantian and
   Kumar, Varun and
   Fulton, Nathan and
   Farahani, Arash and
   Jain, Siddhartha and
   Giaquinto, Robert and
   Qian, Haifeng and
   Ramanathan, Murali Krishna and
   Nallapati, Ramesh and
   Ray, Baishakhi and
   Bhatia, Parminder and
   Sengupta, Sudipta and
   Roth, Dan and
   Xiang, Bing},
  doi = {10.48550/ARXIV.2210.14868},
  url = {https://arxiv.org/abs/2210.14868},
  keywords = {Machine Learning (cs.LG), Computation and Language (cs.CL), FOS: Computer and information sciences, FOS: Computer and information sciences},
  publisher = {arXiv},
  year = {2022},
  copyright = {Creative Commons Attribution 4.0 International}
}

Contributions

skgouda@ benathi@