Datasets:
348 tasks share a PR and base commit with a Multi-SWE-bench, SWE-PolyBench or SWE-bench Multilingual task (pair lists + filtered IDs)
Hi, and thanks for releasing SWE-rebench-V2. I checked it against three multilingual SWE benchmarks. 348 of V2's 32,079 tasks (1.1%) are the same pull request at the same base commit as a benchmark task. People who train on V2 and report these benchmarks may want to filter them.
- Multi-SWE-bench: 83 of 1,632 tasks (5.1%), plus 14 of the 105 Kotlin tasks added to its repo in July 2026, after V2's release.
- SWE-PolyBench: 239 of 2,110 (11.3%), including 41 of the 382 in its Verified subset. 202 of the 239 are
serverless/serverless. - SWE-bench Multilingual: 12 of 300 (4.0%).
- SWE-rebench-V2-PRs: 20 more, matched by repo and PR number: 3 in Multi-SWE-bench, 6 in SWE-PolyBench and 11 in Multilingual. 15 of the 20 are at the same base commit.
Matching takes some care. Multilingual stores repo names in lowercase, and V2 lists some clap and prettier PRs under the repos' old names (kbknapp/clap-rs, jlongster/prettier), so an exact (repo, PR) join misses some. Matching on (base commit, PR number) finds all 348. A 13-gram filter on problem statements would also flag 347 of the 348. For V2-PRs, whose statements are LLM-written, none of the 20 shares a 13-gram with its benchmark task.
What this does not show: that V2's builders chose benchmark PRs or broke a rule they stated. V2 doesn't claim to exclude these benchmarks.
- For Multi-SWE-bench I see no sign of selection either way. Its PRs are in V2 at about the same rate as the PRs its own pipeline discarded: 83 vs 77.4 expected (sd 4.1; z = 1.4; one-sided p = 0.11 for an excess and 0.93 for a shortfall).
- That test has little power, so it would only catch heavy filtering. I have no such test for the other two benchmarks.
- Matching is PR-level only, so the same issue fixed in a different PR is not counted.
- It also says nothing about any model's scores.
The script (pinned revisions), the pair lists and a list of the 31,731 V2 tasks not matched to these benchmarks are here: https://github.com/raimondasl/agentleak#case-3-swe-rebench-v2-vs-multilingual-swe-benchmarks
One question, if you have a moment. None of SWE-bench's 12 test repos is in V2; the license and linked-issue rules may explain pylint and django. Meanwhile, each multilingual benchmark shares a quarter or more of its repos with V2. Was V2 filtered against SWE-bench or any other benchmark, and if so, how: by repo, by PR or by text? That would tell users what is already covered and what they still need to filter.
Would a note on the dataset card help? I'm happy to open a PR with the ID list.
Thanks for the detailed analysis and for sharing the overlap lists!
Yes, we excluded all repositories from the original SWE-bench when building SWE-rebench-V2. We filtered by repository. We did not explicitly filter against the other benchmarks you listed.
We will update the dataset card to explain this, add a warning about the overlaps, and link to your analysis and ID lists so users can filter these tasks before training. Thanks for pointing this out!
Thank you, that's very helpful, and thanks for adding the note to the card. I've added your explanation of the SWE-bench exclusion to the README.
For the link, these lists are pinned to V2 revision 475dd5e8, so the IDs may need a rerun if the dataset changes:
overlap_v2.csv: the 348 V2 tasks, each with its matching benchmark instance;v2_not_in_benchmarks.txt: the other 31,731 V2 tasks;overlap_v2_prs.csv: the 20 V2-PRs matches by repo and PR.
One more thing users may want to drop: 3 V2 tasks have word-for-word the same issue text as a benchmark task but a different PR. These are valkey-io/valkey #809, zeromicro/go-zero #2041 and briannesbitt/Carbon #2667, listed under text_check.same_issue_text_different_pr in summary.json. The PR-level lists above don't include them.
If you release a filtered revision, I'm happy to rerun the check against it.