IssueDupes

A huge dataset of duplicated code issue pairs from Github. Over 1.2M issues, composing over 750K primary-duplicate issue pairs.

Description

This dataset contains pairs of code issues raised in open-source repositories on Github, where one issue was identified to be the duplicate of another. For both the primary issue and the duplicate, the title and body are provided. The dataset spans 25,000 of the most popular open-source repositories, containing over 1.2M issues and over 250M words.

This dataset is currently invite only, intended for research and enterprise uses, but a smaller open dataset is also available - IssueDupesLite

Uses

• Code duplication/similarity detection
• Code embeddings
• Semantic understanding

Dataset Structure

Each row represents a pair of issues, where the 'duplicate' issue has been identified as a duplicate of the 'primary' issue.

Columns are defined as follows:

• name - the full unique repository name (owner/repo)
• owner - the owner of the dataset in Github
• repo - the repository name in Github
• primary_issue_num - the Github issue number of the primary issue
• primary_title - the title of the primary issue
• primary_body - the text body of the primary issue description
• primary_issue_created_at - the creation time of the primary issue (note: this can sometimes be after duplicate_issue_created_at if this issue was deemed more important and the other one closed as a duplicate)
• duplicate_issue_num - the Github issue number of the duplicate issue
• duplicate_title - the title of the duplicate issue
• duplicate_body - the text body of the duplicate issue description
• duplicate_issue_created_at - the creation time of the duplicate issue
• duplicate_closing_comment - the final comment of the duplicate issue, which has often been used to identify the duplication
• comment_created_at - the time of the comment that identified the duplication

Bias, Risks, and Limitations

Duplication comes from the manual labels of repository owners and contributors on Github. While it is generally high quality, there will be some human error. Also, not every true duplicate issue is closed as such, some are just closed without identifying the primary issue for speed. So there is a small chance that there are 'true duplicate' issues on different rows of the dataset that have not been labelled as such.

More Information

For more, see Primordic