BenchAtlas

Sources / swebench

SWE-bench Leaderboard

healthydefault: community

www.swebench.com · trust rank 80 · adapter swebench

Observations
365
append-only score facts
Raw snapshots
8
immutable fetched payloads
Last success
2026-07-14 00:15Z
end of last succeeded/partial run
Freshness
1.4 h
hours since last success

License & attribution

The swe-bench/experiments leaderboard data has no LICENSE file — reused as facts with attribution and a link back (the main SWE-bench repo is MIT).

What we ingest

Embedded leaderboard JSON from each split page (Verified, Lite, bash-only, full, Multimodal, Multilingual) — every entry is an agent+model system. Maintainer-checked entries land as community-verified; unchecked ones as unverified. bash-only standardizes on mini-swe-agent, our canonical harness for composite purposes.

Ingestion runs

last 3 of the run history
StartedDurationStatus
2026-07-14 00:15Z4.7 ssucceeded
2026-07-14 00:09Z5.1 ssucceeded
2026-07-13 20:03Z4.1 spartial

partial = run finished but some rows are blocked on identity review — nothing is published until every referenced model/agent/benchmark is resolved.

Raw snapshots

latest 5
FetchedURLSHA-256
2026-07-14 00:15Zhttps://www.swebench.com/multimodal.html7d6799e3ea3a
2026-07-14 00:09Zhttps://www.swebench.com/multimodal.html6e8e834c5754
2026-07-13 20:03Zhttps://www.swebench.com/original.htmldc0b44886a4e
2026-07-13 20:03Zhttps://www.swebench.com/multilingual-leaderboard.html01e08f05371b
2026-07-13 20:03Zhttps://www.swebench.com/multimodal.htmlb7dafdd1b407

Snapshots are immutable raw payloads — every published number traces back to one. Re-fetching an identical payload bumps the re-fetch time instead of storing a duplicate. Scores from this source appear on benchmark profiles and model pages with this source linked.