Datasets:

Devy1
/

humaneval-no-resource

pretty_name: HumanEval-no-resource
license: mit
language:
  - en
task_categories:
  - text-generation
tags:
  - benchmark
  - evaluation
  - multilingual
  - code-generation
  - low-resource
  - no-resource
dataset_info:
  features:
    - name: task_id
      dtype: int64
    - name: original_humaneval_id
      dtype: string
    - name: prompt
      dtype: string
    - name: tests
      dtype: string
  splits:
    - name: gleam
      num_bytes: 372702
      num_examples: 154
    - name: moonbit
      num_bytes: 133847
      num_examples: 154
  download_size: 158902
  dataset_size: 506549
configs:
  - config_name: default
    data_files:
      - split: gleam
        path: data/gleam-*
      - split: moonbit
        path: data/moonbit-*

HumanEval Benchmark for Gleam and MoonBit

Dataset Description

This dataset contains a translation of the HumanEval benchmark into two no-resource programming languages: Gleam and MoonBit.

This dataset is derived from the MultiPL-E benchmark, which extends HumanEval by translating its programming tasks into multiple programming languages.

More details on this benchmark and how it was created can be found in the paper No Resource, No Benchmarks, No Problem? Evaluating and Improving LLMs for Code Generation in No-Resource Languages.

Dataset Statistics

Number of tasks: 154
Programming languages: 2
Total examples: 308
Source benchmark: HumanEval
Intermediate benchmark: MultiPL-E

Supported Languages

Gleam
MoonBit

Dataset Structure

Each example contains the following fields:

Field	Description
`task_id`	Identifier of the translated task in this benchmark.
`original_humaneval_id`	Identifier of the task in the original HumanEval benchmark.
`prompt`	Signature and docstring of the function to be implemented.
`tests`	Test suite used to evaluate the correctness of generated solutions.

Usage

To load this benchmark, you can use the datasets library as follows.

First, install the library if you haven't already:

pip install datasets

Then, load the dataset:

from datasets import load_dataset

dataset = load_dataset("Devy1/humaneval-no-resource")

# Load the Gleam split
gleam_split = dataset["gleam"]
print(gleam_split[0])

Or load only a specific language:

gleam_ds = load_dataset("Devy1/humaneval-no-resource", split="gleam")

Note: Executing the test suites requires the corresponding language environments and dependencies. The complete generation and evaluation pipeline, including setup instructions for Gleam and MoonBit, is available in our GitHub repository.

Citation

If you use this dataset, please cite our work:

@article{giagnorio2026noresource,
  title={No Resource, No Benchmarks, No Problem? Evaluating and Improving LLMs for Code Generation in No-Resource Languages},
  author={Giagnorio, Alessandro and Martin-Lopez, Alberto and Bavota, Gabriele},
  journal={IEEE Transactions on Software Engineering},
  year={2026},
  publisher={IEEE}
}

Since this dataset is derived from HumanEval through MultiPL-E, please also cite the original benchmark papers:

@article{chen2021evaluating,
  title={Evaluating large language models trained on code},
  author={Chen, Mark and Tworek, Jerry and Jun, Heewoo and Yuan, Qiming and Pinto, Henrique Ponde De Oliveira and Kaplan, Jared and Edwards, Harri and Burda, Yuri and Joseph, Nicholas and Brockman, Greg and others},
  journal={arXiv preprint arXiv:2107.03374},
  year={2021}
}

@article{cassano2023multipl,
  title={Multipl-e: A scalable and polyglot approach to benchmarking neural code generation},
  author={Cassano, Federico and Gouwar, John and Nguyen, Daniel and Nguyen, Sydney and Phipps-Costin, Luna and Pinckney, Donald and Yee, Ming-Ho and Zi, Yangtian and Anderson, Carolyn Jane and Feldman, Molly Q and others},
  journal={IEEE Transactions on Software Engineering},
  volume={49},
  number={7},
  pages={3675--3691},
  year={2023},
  publisher={IEEE}
}

Paper

Hugging Face Papers: https://huggingface.co/papers/2606.16827