metadata

annotations_creators:
  - expert-generated
language:
  - code
language_creators:
  - found
license:
  - mit
multilinguality:
  - monolingual
pretty_name: codequeries
size_categories:
  - 100K<n<1M
source_datasets:
  - original
tags:
  - code
  - code question answering
  - code semantic parsing
  - codeqa
task_categories:
  - question-answering
task_ids:
  - extractive-qa

Dataset Card for Codequeries

Table of Contents
Dataset Description
Dataset Structure
- How to use
- Data Splits and Data Fields
Dataset Creation
Additional Information
- Licensing Information
- Citation Information

Dataset Description

Homepage: Codequeires
Repository: Code repo
Paper:

Dataset Summary

CodeQueries allows to explore extractive question-answering methodology over code by providing semantic natural language queries as question and code spans as answer or supporting fact. Given a query, finding the answer/supporting fact spans in code context involves analysis complex concepts and long chains of reasoning. The dataset is provided with five separate settings; details on the setting can be found in the paper.

Supported Tasks and Leaderboards

Query comprehension for code, Extractive question answering for code.

Languages

The dataset contains code context from python files.

Dataset Structure

How to use

The dataset can directly used with huggingface datasets. You can load and iterate through the dataset for the proposed five settings with the following two lines of code:

import datasets

# instead `twostep`, other settings are <ideal/file_ideal/prefix>.
ds = datasets.load_dataset("thepurpleowl/codequeries", "twostep", split=datasets.Split.TEST)
print(next(iter(ds)))

#OUTPUT:
{'query_name': 'Unused import',
 'code_file_path': 'rcbops/glance-buildpackage/glance/tests/unit/test_db.py',
 'context_block': {'content': '# vim: tabstop=4 shiftwidth=4 softtabstop=4\n\n# Copyright 2010-2011 OpenStack, LLC\ ...',
                    'metadata': 'root',
                    'header': "['module', '___EOS___']",
                    'index': 0},
 'answer_spans': [{'span': 'from glance.common import context',
                   'start_line': 19,
                   'start_column': 0,
                   'end_line': 19,
                   'end_column': 33}
                 ],
 'supporting_fact_spans': [],
 'example_type': 1,
 'single_hop': False,
 'subtokenized_input_sequence': ['[CLS]_', 'Un', 'used_', 'import_', '[SEP]_', 'module_', '\\u\\u\\uEOS\\u\\u\\u_', '#', ' ', 'vim', ':', ...],
 'label_sequence': [4, 4, 4, 4, 4, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, ...],
 'relevance_label': 1
}

Data Splits and Data Fields

Detailed information on the data splits for proposed settings can be found in the paper.

In general, data splits in all prpoposed settings have examples in following fields -

  - query_name (query name to uniquely identify the query)
  - code_file_path (relative source file path w.r.t. ETH Py150 corpus)
  - context_blocks (code blocks as context with metadata)  [`prefix` setting doesn't have this field and `twostep` has `context_block`]
  - answer_spans (answer spans with metadata)
  - supporting_fact_spans (supporting-fact spans with metadata)
  - example_type (1(positive)) or 0(negative)) example type)
  - single_hop (True or False - for query type)
  - subtokenized_input_sequence (example subtokens)  [`prefix` setting has the corresponding token ids]
  - label_sequence (example subtoken labels)
  - relevance_label (0 (not relevant) or 1 (relevant) - relevance label of a block)   [only `twostep` setting has this field]

Dataset Creation

The dataset is created by using ETH Py150 Open corpus as source for code contexts. To get natural language queries and corresponding answer/supporting spans in ETH Py150 Open corpus files, CodeQL was used.

Dataset Card for Codequeries

Table of Contents