Backporting Battles on the Branches: Agentic vs Human Fixes

Replication package for the empirical study comparing LLM-signaled agentic backport pull requests (PRs) with human-authored backport PRs on GitHub stable branches. This dataset supports Backporting Battles on the Branches: Agentic vs Human Fixes. It contains enriched PR metadata, file-level diffs, and review/discussion events collected from a popularity-ranked Top-3500 GitHub repository universe through a two-phase discovery-and-fetch pipeline, followed by backport confirmation, enrichment, and cohort balancing. Cohort sizes Enriched release: 1,502 confirmed agentic backports (622 LLM-signaled, 880 automation-only) and 1,502 confirmed human backports from overlapping repositories. Primary analysis subsample: 622 LLM-signaled agentic vs 622 human PRs (balanced, seed 42). Each agentic row includes agent, agent_kind (llm or automation), agent_confidence, and agent_signal. Contents dataset/ top3500_software_repos.csv Star-ranked repository list top3500/repos_with_agentic.json Repositories passing Phase-1 discovery top3500/enriched/ prs_backport_agentic.parquet Confirmed agentic backport PR rows prs_backport_human.parquet Confirmed human backport PR rows file_changes_backport.parquet Per-file diff summaries reviews_backport.parquet Formal PR reviews + issue comments failed_file_change_fetches.csv Log of file-diff fetch failures validation/ backport_spotcheck_labelsDone.csv Backport-ID spot-check (120 PRs, adjudicated) lexicon_validation_sampleDone.csv Comment semantics validation (200 comments) failure_reason_validation_sampleDone.csv Unmerged failure-reason validation (120 PRs) Citation If you use this dataset, please cite this Zenodo record (DOI: 10.5281/zenodo.21048815) and the accompanying paper. License: CC-BY 4.0