Hugging Face's logo

Datasets:
BeNatural
/
github-scikit-learn-issues-dataset

Tasks:
Text Classification
Sentence Similarity
Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
Size:
< 1K
ArXiv:
Tags:
github
scikit-learn
issue-tracking
software-engineering
open-data
text
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
github-scikit-learn-issues-dataset / README.md
BeNatural's picture
BeNatural
Update README.md
00aa5a6 verified
preview code
|
Raw History Blame Contribute Delete
9.92 kB
metadata 
pretty_name: GitHub Scikit-learn Issues Dataset
language:
  - en
license: other
size_categories:
  - n<1K
task_categories:
  - text-classification
  - sentence-similarity
tags:
  - github
  - scikit-learn
  - issue-tracking
  - software-engineering
  - open-data
  - text

GitHub Scikit-learn Issues Dataset

1. Overview

This dataset contains 300 issues collected from the public GitHub repository scikit-learn/scikit-learn. It was created for an AI/data assessment (Weeks 9-12) to demonstrate dataset creation, a Transformers pipeline and FAISS semantic search, with documentation of sourcing, limitations and compliance.

2. Why this dataset

GitHub issues are real, human-written technical text (bug reports, feature requests, documentation problems) paired with structured metadata (labels, dates, state, comment counts). This makes them suitable for text classification, sentiment/zero-shot analysis and semantic search for similar issues, and the scikit-learn project is well known, openly developed and well documented.

3. Data collection

  • Platform: GitHub (public issues of scikit-learn/scikit-learn)
  • Collection method: issues were retrieved from the GitHub REST API (https://api.github.com/repos/scikit-learn/scikit-learn/issues) using Python and the requests library in a Google Colab notebook. Requests used state=all, 100 records per page, sorted by creation date in ascending order (oldest first), with a short pause between pages. Pages were collected until 300 genuine issues had been gathered, and the result was trimmed to exactly 300. The records were then assembled into a pandas DataFrame, saved as CSV and Parquet, and published to the Hugging Face Hub.
  • Collection date: 12 October 2026. The collection notebook was re-run on 5 November 2026 and returned the same 300 issues and the same summary figures.
  • Authentication: requests were made to the public GitHub API without a personal access token, within GitHub's public rate limits. No credentials were used or stored.
  • Scope: the dataset contains issues only. The GitHub issues endpoint also returns pull requests, so any record containing a pull_request field was removed (about half of the early records returned were pull requests). No record in the dataset has a /pull/ URL.
  • Selection: because results were sorted oldest first, the 300 records are the oldest issues in the repository (issue numbers 1 to 768, created between 2010-08-31 and 2012-04-13). They are not a random sample.

4. Dataset structure

Column Type Description
issue_id int GitHub's global issue ID
issue_number int Issue number within the repository
repository string scikit-learn/scikit-learn
title string Issue title
body string Issue description (9 records are empty)
state string closed (295) or open (5)
labels string Comma-separated GitHub labels (empty for 235 records)
created_at datetime When the issue was opened (UTC)
closed_at datetime When the issue was closed (empty for open issues)
comments int Number of comments on the issue (count only; comment text is not included)
author string Public GitHub username of the issue author
html_url string Link to the original issue
text_length int Derived field: length of the issue text (see Section 5)
time_to_close_hours float Derived field: hours between created_at and closed_at (empty for open issues)

5. Processing and augmentation

  • Raw issue records were assembled into a structured table (one row per issue), keeping 14 selected fields from the many returned by the API.
  • Pull requests were excluded so that only issues remain.
  • Text cleaning: leading and trailing whitespace was removed from titles and bodies, a missing body was stored as an empty text, and the list of labels was joined into one comma-separated text field.
  • Duplicates: records were de-duplicated on issue_id and records with empty titles were removed. No records needed removing (all issue_id, issue_number and html_url values are unique).
  • Validation: the notebook checks that issue IDs are unique, no title is empty, and the row count is in the expected range.
  • Derived fields added: text_length and time_to_close_hours.
  • Missing values were kept, not imputed: 9 empty bodies, 235 records without labels, and 5 open issues without closed_at / time_to_close_hours.

6. Initial data summary

Item Value
Rows / columns 300 / 14
Date range 2010-08-31 to 2012-04-13
State 295 closed, 5 open
Issues with labels 65 of 300 (22%)
Most common labels New Feature (21), Enhancement (11), Bug (7)
Comments per issue median 4, mean 7.2, max 59
Time to close median about 1,195 hours (about 50 days); highly skewed (max about 125,700 hours)
Distinct authors 84; the top 3 (amueller, fabianp, mblondel) wrote 132 of the 300 issues

7. Intended uses

  1. Transformers pipeline: apply a pretrained model (e.g. sentiment or zero-shot classification) to titles/bodies and examine the outputs.
  2. Semantic search: embed issue text and build a FAISS index to retrieve similar issues for natural-language queries.

Not intended for: profiling individual contributors, or drawing conclusions about the modern scikit-learn project.

8. Limitations and bias

  • Temporal bias: all issues are from 2010-2012, the earliest days of the project. Topics, labels and writing style differ from current issues, and many early issues were written by a handful of core developers.
  • Small sample: 300 rows, not representative of the repository's full history.
  • Author concentration: a few authors account for a large share of records.
  • Sparse labels: 78% of issues have no label, so label-based analysis covers a minority.
  • Short or empty text: some bodies are empty or very short, which weakens classification and embedding quality.
  • Open vs closed: only 5 issues are open, so resolution-time analysis is dominated by closed issues.
  • Model limitations: general-purpose sentiment models often misread technical text; semantic search quality depends on the embedding model.
  • No comment text: only the number of comments is included.

9. Ethics, legal and compliance

  • Privacy: issues are public, but they include public GitHub usernames and can contain links or occasional email-like text. Usernames have been kept as they appear publicly on GitHub. The dataset is not used to identify or profile individuals.
  • Attribution: content originates from the scikit-learn GitHub project and its contributors; every record links back to its original issue via html_url.
  • Licensing: the scikit-learn software is BSD-3-Clause. This licence is for the software repository and is not assumed to cover user-written issue text, so this dataset is marked license: other and is shared for educational and research purposes only. Users should check GitHub's Terms of Service and the original project's terms before reuse.
  • Terms of service: data was collected from publicly available content through the official GitHub API.
  • Organisational policy alignment: the dataset was collected, stored and documented in line with the organisation's data storage, management and compliance policies provided for this assessment: public source data only, no private data, attribution to the original source, no credentials needed or stored, and versioned storage.
  • Security: no API token or other credential was used, so none is stored in the notebooks or this repository.

10. Risks and corrective actions

Risk Mitigation
Personal data in public issue text Keep only public handles; document handling; remove or anonymise author if required; do not target individuals
Temporal / sampling bias Document the 2010-2012 window; in future, collect a random or time-stratified sample
Licence ambiguity of user content Mark as other, link back to originals, educational use only
Missing labels / bodies Keep missing values visible, report counts, avoid over-interpreting
Misleading model output Manually review sample predictions and state model limits
Data loss Dataset on the Hugging Face Hub with version history; notebooks and outputs backed up per policy

11. Storage, versioning and reproducibility

  • Published on the Hugging Face Hub (git-based, so every commit is versioned): https://huggingface.co/datasets/BeNatural/github-scikit-learn-issues-dataset
  • Notebooks: three notebooks (data collection and augmentation, Transformers pipeline, FAISS semantic search) are submitted with this assessment as evidence.
  • Local copy: the CSV and notebooks are kept as a backup copy alongside the project files.
  • Version: v1.0, collected 12 October 2026 and re-verified on 5 November 2026. Any later changes to the data or processing should be recorded here.

12. References