MultiPL-E HumanEval

Overview

MultiPL-E HumanEval is a multilingual code generation benchmark derived from OpenAI's HumanEval. It extends the original HumanEval to 18 programming languages, enabling cross-lingual evaluation of code generation capabilities.

Task Description

Task Type: Multilingual Code Generation
Input: Programming problem prompt with function signature and docstring
Output: Complete function implementation that passes test cases
Languages: 18 languages (C++, TypeScript, Shell, C#, Go, Java, Lua, JavaScript, PHP, Perl, Racket, R, Rust, Scala, Swift, Ruby, D, Julia)

Key Features

Multilingual evaluation across 18 programming languages
Execution-based evaluation with test cases
Supports pass@k metric for code generation
Docker sandbox environment for safe code execution
Derived from HumanEval with consistent problem difficulty

Evaluation Notes

Sandbox Required: Requires sandbox environment for safe code execution
Default evaluation uses test split
Primary metric: Accuracy with pass@k aggregation
Timeout: 30 seconds per test case
See sandbox documentation for setup

Properties

Property	Value
Benchmark Name	`multiple_humaneval`
Dataset ID	evalscope/MultiPL-E
Paper	N/A
Tags	`Coding`
Metrics	`acc`
Default Shots	0-shot
Evaluation Split	`test`
Aggregation	`mean_and_pass_at_k`

Data Statistics

Metric	Value
Total Samples	2,864
Prompt Length (Mean)	696.59 chars
Prompt Length (Min/Max)	285 / 2463 chars

Per-Subset Statistics:

Subset	Samples	Prompt Mean	Prompt Min	Prompt Max
`humaneval-cpp`	161	797.74	351	2057
`humaneval-ts`	159	639.27	320	1529
`humaneval-sh`	158	648.14	325	1534
`humaneval-cs`	158	983.67	538	2335
`humaneval-go`	154	701.14	351	1616
`humaneval-java`	158	1009.53	531	2463
`humaneval-lua`	161	620.31	294	1497
`humaneval-js`	161	611.34	287	1490
`humaneval-php`	161	632.01	298	1551
`humaneval-pl`	161	611.27	296	1481
`humaneval-rkt`	161	634.22	304	1537
`humaneval-r`	161	607.78	285	1484
`humaneval-rs`	156	686.67	313	1591
`humaneval-scala`	160	832.81	423	1998
`humaneval-swift`	158	660.88	323	1556
`humaneval-rb`	161	611.55	289	1474
`humaneval-d`	156	640.72	328	1501
`humaneval-jl`	159	616.4	303	1478

Sample Example

Subset: humaneval-cpp

{
  "input": [
    {
      "id": "8f096947",
      "content": "```cpp\n#include<assert.h>\n#include<bits/stdc++.h>\n// Check if in given vector of numbers, are any two numbers closer to each other than\n// given threshold.\n// >>> has_close_elements((std::vector<float>({(float)1.0f, (float)2.0f, (float)3.0f}) ... [TRUNCATED] ... ents(std::vector<float> numbers, float threshold) {\n\n```\n\nPlease complete the above code according to the requirements in the docstring. Write the complete code and wrap it in markdown fenced code. The code should not contain `Main` function."
    }
  ],
  "target": "",
  "id": 0,
  "group_id": 0,
  "metadata": {
    "tests": "}\nint main() {\n    auto candidate = has_close_elements;\n    assert(candidate((std::vector<float>({(float)1.0f, (float)2.0f, (float)3.9f, (float)4.0f, (float)5.0f, (float)2.2f})), (0.3f)) == (true));\n    assert(candidate((std::vector<float>({( ... [TRUNCATED] ... (std::vector<float>({(float)1.1f, (float)2.2f, (float)3.1f, (float)4.1f, (float)5.1f})), (1.0f)) == (true));\n    assert(candidate((std::vector<float>({(float)1.1f, (float)2.2f, (float)3.1f, (float)4.1f, (float)5.1f})), (0.5f)) == (false));\n}\n",
    "stop_tokens": [
      "\n}"
    ],
    "task_id": "HumanEval_0_has_close_elements",
    "language": "cpp",
    "doctests": "transform"
  }
}

Note: Some content was truncated for display.

Prompt Template

Prompt Template:

{prompt}

Sandbox Configuration

This benchmark requires a sandbox environment for code execution.

{
  "image": "volcengine/sandbox-fusion:server-20250609",
  "tools_config": {
    "shell_executor": {},
    "python_executor": {},
    "multi_code_executor": {}
  },
  "memory_limit": "2g",
  "cpu_limit": "2.0"
}

Usage

Using CLI

evalscope eval \
    --model YOUR_MODEL \
    --api-url OPENAI_API_COMPAT_URL \
    --api-key EMPTY_TOKEN \
    --datasets multiple_humaneval \
    --sandbox '{"enabled": true}' \
    --limit 10  # Remove this line for formal evaluation

Using Python

from evalscope import run_task
from evalscope.config import TaskConfig

task_cfg = TaskConfig(
    model='YOUR_MODEL',
    api_url='OPENAI_API_COMPAT_URL',
    api_key='EMPTY_TOKEN',
    datasets=['multiple_humaneval'],
    sandbox={'enabled': True},
    dataset_args={
        'multiple_humaneval': {
            # subset_list: ['humaneval-cpp', 'humaneval-ts', 'humaneval-sh']  # optional, evaluate specific subsets
        }
    },
    limit=10,  # Remove this line for formal evaluation
)

run_task(task_cfg=task_cfg)

5.3 KiB Raw Blame History