M2CRB

Dataset Summary

M2CRB contains pairs of text and code data with multiple natural and programming language pairs. Namely: Spanish, Portuguese, German, and French, each paired with code snippets for: Python, Java, and JavaScript. The data is curated via an automated filtering pipeline from source files within The Stack followed by human verification to ensure accurate language classification I.e., humans were asked to filter out data for which natural language did not correspond to a target language.

Supported Tasks

M2CRB is a multilingual evaluation dataset for code-to-text and/or text-to-code models, both on information retrieval or conditional generation evaluations.

Currently Supported Languages

NATURAL_LANGUAGE_SET = {"es", "fr", "pt", "de"}
PROGRAMMING_LANGUAGE_SET = {"python", "java", "javascript"}

How to get the data with a given language combination

from datasets import load_dataset

def get_dataset(prog_lang, nat_lang):
    test_data = load_dataset("blindsubmissions/M2CRB")

    test_data = test_data.filter(
        lambda example: example["docstring_language"] == nat_lang
        and example["language"] == prog_lang
    )

    return test_data

Dataset Structure

Data Instances

Each data instance corresponds to function/methods occurring in licensed files that compose The Stack. That is, files with permissive licences collected from GitHub.

Relevant Data Fields

• identifier (string): Function/method name.
• parameters (string): Function parameters.
• return_statement (string): Return statement if found during parsing.
• docstring (string): Complete docstring content.
• docstring_summary (string): Summary/processed docstring dropping args and return statements.
• function (string): Actual function/method content.
• argument_list (null): List of arguments.
• language (string): Programming language of the function.
• docstring_language (string): Natural language of the docstring.
• type (string): Return type if found during parsing.

Summary of data curation pipeline

• Filtering out repositories that appear in CodeSearchNet.
• Filtering the files that belong to the programming languages of interest.
• Pre-filtering the files that likely contain text in the natural languages of interest.
• AST parsing with Tree-sitter.
• Perform language identification of docstrings in the resulting set of functions/methods.
• Perform human verification/validation of the underlying language of docstrings.

Social Impact of the dataset

M2CRB is released with the aim to increase the coverage of the NLP for code research community by providing data from scarce combinations of languages. We expect this data to help enable more accurate information retrieval systems and text-to-code or code-to-text summarization on languages other than English.

As a subset of The Stack, this dataset inherits de-risking efforts carried out when that dataset was built, though we highlight risks exist and malicious use of the data could exist such as, for instance, to aid on creation of malicious code. We highlight however that this is a risk shared by any code dataset made openly available.

Moreover, we remark that while unlikely due to human filtering, the data may contain harmful or offensive language, which could be learned by the models.

Discussion of Biases

The data is collected from GitHub and naturally occurring text on that platform. As a consequence, certain language combinations are more or less likely to contain well documented code and, as such, resulting data will not be uniformly represented in terms of their natural and programing languages.

Known limitations

While we cover 16 scarce combinations of programming and natural languages, our evaluation dataset can be expanded to further improve its coverage. Moreover, we use text naturally occurring as comments or docstrings as opposed to human annotators. As such, resulting data will have high variance in terms of quality and depending on practices of sub-communities of software developers. However, we remark that the task our evaluation dataset defines is reflective of what searching on a real codebase would look like. Finally, we note that some imbalance on data is observed due to the same reason since certain language combinations are more or less likely to contain well documented code.

Maintenance plan:

The data will be kept up to date by following The Stack releases. We should rerun our pipeline for every new release and add non-overlapping new content to both training and testing partitions of our data.

This is so that we carry over opt-out updates and include fresh repos.

Update plan:

• Short term:

  • Cover all 6 programming languages from CodeSearchNet.

• Long-term

  • Add an extra test set containing human-generated text/code pairs so the gap between in-the-wild and controlled performances can be measured.
  • Include extra natural languages.

Licensing Information

M2CRB is a subset filtered and pre-processed from The Stack, a collection of source code from repositories with various licenses. Any use of all or part of the code gathered in M2CRB must abide by the terms of the original licenses.