JetBrains-Research/template-generation

Template Generation Dataset for AI Agents Evaluation

This is the benchmark for the Project Template Generation task, which requires creating a project template (content and structure) by short textual description.

The dataset provides all the required components for evaluation of project template generation approaches in real project templates collected from GitHub, including:

• Repository description;
• Repository root README.md file content;
• Repository link by which the "golden" template can be accessed;
• GitHub repository telemetry, including additional data and metrics that can be useful in developing new approaches.

All the repositories are published under permissive licenses (MIT, Apache-2.0, BSD-3-Clause, and BSD-2-Clause). The datapoints can be removed upon request.

The collected dataset was carefully filtered, enhanced with useful metrics and, what's more, manually labeled, which assures the data quality and provides a golden subset of good examples for evaluation.
Moreover, the dataset was split into several categories, namely:

Category	Description	Number of data points
py	Repositories with Python main language	565
java	Repositories with Java main language	81
kt	Repositories with Kotlin main language	19
android	Repositories with Kotlin/Java main language based on Android sdk	17

...and splits, namely:

Split	Description
dev	All collected datapoints
test	Manually verified datapoints
train	Rest of the datapoint from dev without test

The following sections describe the utilities around the dataset, as well as dataset content.

Dataset Collection

This dataset contains information about repos (initially gathered from https://seart-ghs.si.usi.ch) matching the following criteria:

• Python, Java, Kotlin programming languages
• 10+ stars
• 10-1000 code lines
• updated after 2023-01-01 00:00
• not forks
• permissive licenses (MIT License, Apache License 2.0, BSD 3-Clause "New" or "Revised" License, BSD 2-Clause "Simplified" License)
• filtered by is_template=True or template-related keywords presence in the description (template, boilerplate, starter, skeleton, blueprint, scaffold, pattern, seed, example, demo, sample, showcase, illustration, exemplar, use case, prototype)
• android is moved to separate category (by android keyword in description or repo fill_name)

You can find all scripts to reproduce dataset collection in our GitHub repository

Dataset Description

Field	Description
id	Identifier of data point.
full_name	Repository full name {owner}/{name}.
owner	Bug issue repository owner.
name	Bug issue repository name.
html_url	GitHub link to issue https://github.com/{owner}/{name}.
is_template	True if the repositories marked as a template, otherwise False.
description	Repository description.
template_keywords	Template-related keywords.
license	Repository license (one from 'MIT License', 'Apache License 2.0', 'BSD 3-Clause "New" or "Revised" License', 'BSD 2-Clause "Simplified" License').
topics	Repository topics.
size	Repo size [MB].
metrics	Dict from languages to their meta info like lines count.
languages	Repo languages.
language	Repo main language.
created_at	Date of the repo was created in format of yyyy-mm-dd hh:mm:ss.
updated_at	Date of the last repo update in format of yyyy-mm-dd hh:mm:ss.
code_lines	Number of lines of code in repo.
repo_symbols_count	Number of symbols in the repository.
repo_tokens_count	Number of tokens in the repository.
repo_words_count	Number of words in the repository.
repo_lines_count	Number of lines in the repository.
repo_files_count	Number of files in the repository.
repo_code_symbols_count	Number of code symbols in the repository.
repo_code_tokens_count*	Number of code tokens in the repository.
repo_code_words_count	Number of code words in the repository.
repo_code_lines_count	Number of code lines in the repository.
repo_code_files_count	Number of code files in the repository.
description_symbols_count	Number of symbols in the repository description.
description_tokens_count*	Number of tokens in the repository description.
description_words_count	Number of words in the repository description.
description_lines_count	Number of lines in the repository description.
readme	Root README.md repository content.
readme_symbols_count	Number of symbols in the repository readme.
readme_header_tokens_count*	Number of tokens in the repository readme.
readme_header_words_count	Number of words in the repository readme.
readme_header_lines_count	Number of lines in the repository readme.

* Tokens calculated via GPT-4 tokenizer

Dataset analysis

• Topics analysis notebook
• Large-scale analysis notebook

Dataset usage

• Load the data via load_dataset:

from datasets import load_dataset
# Select a configuration from ["java", "kt", "android"]
configuration = "java"
# Select a split from ["dev", "train", "test"]
split = "dev"
# Load data
dataset = load_dataset("JetBrains-Research/template-generation", configuration, split=split)