N = 567 agentic pull requests · 475 merged (83.8%) · 92 rejected
Hover to preview · click Revision or Rejected to drill down · Esc to go back
AIDev-pop is the curated subset of AIDev drawn from repositories with >100 GitHub stars: 33,596 agentic PRs, or 3.60% of the full corpus of 932,791.
Star tier is the paper's sampling boundary, not a quality ranking; shares are computed over PRs. AIDev-pop overlaps with the full corpus, so the counts cannot be added.
Click the donut, the AIDev-pop legend, or the button to view scale, workflow, data structure, and the research agenda.
Five coding agents, 933K pull requests, one collaboration experiment inside real repositories.
AIDev collected 932,791 agentic PRs from GitHub, spanning 116,211 repositories and 72,189 developers, with data cut off on 2025-08-01. It is not a leaderboard or an answer key, but a reusable observational substrate: it joins agent submissions, human reviews, follow-up fixes, and event timelines so researchers can ask real-world questions about adoption, quality, and risk.
The full tables provide scale and coverage; the curated subset preserves the human traces needed for fine-grained analysis. They are a wide-angle view and a close-up of the same data, not two independent corpora.
What it means: The AIDev replication package's README calls the >100 GitHub stars curated subset AIDev-pop (the popular subset), and provides a separate download for its raw data. The paper prose does not use this identifier and instead says “curated subset”; both refer to the same 33,596 PRs, not an additional dataset.
Reading note: percentages are computed against each full-corpus total. The curated and full tables overlap, so their row counts cannot be added, and 3.60% is not an agent adoption rate.
AIDev's structure follows the real GitHub collaboration workflow. The path in Figure 1 explains why the dataset includes PRs, comments, commits, issues, and timelines together.
The agent is assigned an issue in a real project and receives task context and acceptance goals.
It locates files in the repository, changes the implementation, and runs available checks or tests.
It submits a pull request with a title, description, and scope, entering the public collaboration process.
Maintainers comment, request changes, or approve; these judgments become first-hand signals of quality and risk.
The agent responds to comments through follow-up commits and replies, keeping human-agent collaboration moving.
Note: not every coding agent supports handling review comments. "Reviewed" and "review resolved" are therefore different kinds of evidence and should be analyzed separately.
From PR metadata to file-level diffs, reviews, issues, and event streams. Bars compare only within each table group, so records of different magnitudes are not mixed.
all_pull_request932,791title · body · agent · state · timestamps · repository · user
all_repository116,211name · license · language · URL · stars · forks
all_user72,189login · followers · creation date
pull_request33,596Same fields as all_pull_request, limited to the curated subset
repository2,807Same fields as all_repository, limited to the curated subset
user1,796Same fields as all_user, limited to the curated subset
pr_comments39,122Threaded comments: author · body · timestamp
pr_reviews28,875Review verdicts: approve / request changes and metadata
pr_review_comments19,450Inline review: file path · diff hunk · timestamp
pr_commits88,576Commits linked to PRs: SHA · author · message
pr_commit_details711,923File-level diff: additions · deletions · patches
related_issue4,923Mapping between PRs and related issues
issue4,614title · body · state · user · timestamps
pr_timeline325,500Event history: committed · reviewed · labeled · merged · closed
pr_task_type33,596PR purpose classified with GPT assistance by Conventional Commits category (bug fix · feature · docs, etc.)
The 15 tables total 2,415,969 records, but that is not 2.42M independent PRs: comments, commits, issues, and timeline events are different views of the same PR lifecycle. The full and curated tables also overlap.
The paper lays out five research agendas. These are not conclusions built into the dataset, but the next set of empirical questions AIDev is meant to support.
Who uses coding agents? How do adoption paths differ between newcomers and experienced developers? How does penetration vary across repositories and ecosystems?
How do agentic PRs change code: lines added or removed, files touched, and the balance of refactoring and reuse? Do they follow project conventions and CI checks?
How often do agents contribute tests? When tests are missing, do developers add verification through follow-up commits or linked PRs?
Which concerns dominate review: correctness, style, security, or testing? Which comments can agents resolve, and which remain difficult?
How do common defects emerge? Can early signals predict rejection or review cost? Which changes eventually reach production and persist?
Hugging Face supports in-browser SQL through Data Studio; Zenodo provides a citable archive; the GitHub repository includes example notebooks and a Colab entry point.
But acceptance is not the same as a clean pass: 214 of 475 merged PRs (45.1%) were revised before merge.