feat(benchmarks): OpenHands + Terminus2 normalizer recipe — 716 labeled sessions unlocked - #519
Merged
Conversation
… Terminus2 Per-row fault-tolerant preparation for the 742 verified OpenHands and Terminus2 rows: downloads the pinned dataset archives, normalizes through CodeTracer's skill contract, mirrors every traces importer gate before the all-or-nothing bulk import, and emits per-family labels plus a receipt naming every failed row and reason. Upstream CodeTracer (2d302191, repository HEAD) has no parser for the swe_raw completion-log layout, and its openhands/terminus2 seed parsers do not reproduce the annotation's step numbering, so three local skills carry the annotation-faithful conventions with count evidence against every verified row: openhands_completions 313/313, openhands_sessions 183/199, terminus2_commands 220/222. Result: 496/520 OpenHands and 220/222 Terminus2 rows import; the 26 losses are 8 manifest rows without archives, 16 condensed session streams whose published views disagree with the annotated count, and 2 empty archives.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
The prepare-stage recipe that turns the two biggest unimported CodeTraceBench families into analyst-readable instruments:
benchmarks/trace-analysis/codetracebench-oht2-20260801/—prepare-bulk.py+ three CodeTracer user-skills (openhands_completions,openhands_sessions,terminus2_commands), pinned to the same CodeTracer commit (2d302191) and dataset revision (aa213b84) as every prior split.Upstream CodeTracer ships parsers for both families but neither reproduces the annotation's step numbering (Terminus2 episode-level: 11/222 rows match; OpenHands run-events-only: 1/199) and the
swe_rawcompletion-log layout has no parser at any upstream commit — so the numbering conventions were derived empirically against every verified manifest row and enforced fail-loud: a row whose step count cannot be matched is never imported, because a miscounted row silently misaligns every gold label on it.Live result (receipts committed to the bench cache, referenced in the README)
Rule validation across all extracted rows: completions 313/313, sessions 183/199, commands 220/222 step-count matches. Spot proof 6/6 + independent operator re-check: labels == steps.json == OTLP spans.
Why it matters
Every existing split is spent, and split3 was found positionally degenerate post-certification — these 716 never-touched labeled sessions are the only path to honest future certification, and the first cross-agent-family generalization test the analyst has ever had (mini-SWE-only until now).
Companion: tangle-network/traces PR pinning the family import invariants in tests. Gates: biome clean, typecheck exit 0, build green, suite 4656/3 skipped.