Skip to content

docs(results): add RH-SWE-bench Harbor result (baseline 0.58 → best 0… - #313

Merged
OsherElhadad merged 2 commits into
skillberry-ai:mainfrom
AIMikav:docs/rh-swebench-results
Aug 10, 2026
Merged

docs(results): add RH-SWE-bench Harbor result (baseline 0.58 → best 0…#313
OsherElhadad merged 2 commits into
skillberry-ai:mainfrom
AIMikav:docs/rh-swebench-results

Conversation

@AIMikav

@AIMikav AIMikav commented Aug 10, 2026

Copy link
Copy Markdown
Collaborator

What & why

Add RH-SWE-bench (Harbor SWE-bench Verified) result to the results page.

  • Baseline: 0.580 (58.0%) → Best: 0.765 (76.5%) — +18.5pp / +31.9% relative
  • 119 SWE-bench Verified tasks, fit metric (train == val == test)
  • Agent: claude-sonnet-4-6 via Harbor (Docker sandboxed evaluation)
  • Optimizer: claude-code @ claude-opus-4-6, $60.75 over 7 iterations
  • 2 of 7 iterations accepted, 24 tasks improved, 2 regressed

Checklist

  • python -m pytest core/tests -q passes — no code changes, docs only
  • python skills/_registry/build_manifest.py skills is clean — no skill changes
  • Any new/changed skill's scripts/check.py is green — N/A
  • Honesty invariants untouched (gate on val, test sealed) — N/A
  • Docs updated (site/results.html)

AIMikav and others added 2 commits August 10, 2026 10:03
The feat(site) commit 1f0c839 changed benchmarks.html/benchmarks.js so
the f-tier dropdown is now populated at load time from RECORDS data via
hydrateFilter, removing the hardcoded smoke/pilot/full option list.

test_tier_filter_offers_every_tier was written against the old static
approach and failed because the HTML now only carries the 'all' stub
option. Update the test to assert the new invariant: the select element
exists as a stub in the HTML AND benchmarks.js calls hydrateFilter for
'#f-tier', which is what guarantees any tier present in the data is
selectable without requiring a markup change.
@OsherElhadad
OsherElhadad merged commit 1af755b into skillberry-ai:main Aug 10, 2026
10 checks passed
@skillberry-bot

Copy link
Copy Markdown
Contributor

Automatic Labeling Failed

An error occurred while trying to automatically label this pull request. Please check the workflow logs for details and add labels manually.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants