Mining Merged PRs Raises Usable Commit-Untangling Data From 9.5% to 55%, Yielding a 5.7x Larger Dataset

arXiv 2607.26730·low signalXBlueskyLinkedInCopy link

arXiv 2607.26730 (2026-07-29) builds commit-untangling training data cheaply by mining pull requests where the squashed PR is tangled but each individual feature-branch commit is atomic — 'ideal PRs' — which their filtering rules raise from 9.5% to 55% of candidates, producing a dataset more than 5.7x larger than prior heuristic-based ones. They show the PR-derived dataset differs statistically from datasets built with Herzig's heuristics along dimensions that affect confidence voters and are likely to affect learning-based untanglers, and validate that impact on Herzig's original method. A Python dataset built the same way reaches a comparable 56.5% ideal-PR rate.