The dataset card describes the layout (<task>/<run>/..., signature-tasks/<x>/...),
the raw-files rule, how to download one task or clone without large files, how to
contribute (upload with OpenRSI-Index's checks/publish_logs.py, which checks the
logs before uploading them to main or as a Hub pull request; maintainers check a
pull request with publish_logs.py --check-pr N before merging it), and how this
history was converted from OpenRSI-Index. LICENSE is OpenRSI-Index's Apache-2.0
license. .gitattributes also stores common binary file types (images, audio,
arrays, model weights) through LFS, and no longer lists two files that moved under
signature-tasks/ (their new paths have rules of their own).
Task links that pointed into OpenRSI-Index by relative path now use OpenRSI-Index
URLs. The archive READMEs describe the raw agent_output.txt transcripts and their
manifests instead of gzip and chunk restoration, and the SGLang measurement
extractor reads the dataset copy it is stored in (--…
fix(harness): stop recording Agent budget exhaustion as failed
The stop hook blocks voluntary exits, so reaching the Agent timeout is the
normal end of a run. The run outcome now follows the submissions
(completed or no_valid_submission); final_result.json still records
timed_out: true. Real failures (engine errors, infrastructure rounds,
cleanup errors) remain failed.
Drop the rsi-logs README sentences that described the old status.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Publish learnability-cot Codex run with a 7.7 h budget, replacing the short recording
Recorded 2026-09-18 on the OpenRSI-Index 738c0ea task and Harness: codex /
gpt-5.6-sol / xhigh, 27,720 s Agent budget, 27 persisted submissions, best
46.081 (agent-15) against the Judge-measured Mix-Long baseline 46.073. The
run reached its budget, so final_result.json carries the Harness timed-out
marker; every Judge round finished before shutdown.
The previous Codex archive ended after 6,461 s of a 216,000 s budget and did
not reflect an allowance comparable to the Claude recording; it is replaced.
Sanitized per rsi-logs/SANITIZATION.md: run paths and GPU identities in
run-plan.json are anonymized; transcripts, submissions, feedback, verifier
rewards, scores, timings and digests are unchanged. Adds the task-level
README with both runs.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
