sinatras/classical-rl-efficiency-traces

Classical RL efficiency: privacy-cleaned trace dataset

This is a privacy-redacted derivative of the saved research archive, published on Hugging Face. The private original and the September 23 cleaned snapshot remain unchanged. No inference, grading, rescoring or tokenization was run for this release.

License and publication status

Publicly accessible, not open-licensed. The publisher's original protected material is offered under an All Rights Reserved (Proprietary) notice; see LICENSE. Reuse requires permission except where applicable law, Hugging Face's terms, or an applicable third-party license already permits it. Public visibility allows anyone to view or download the repository; the license is not an access-control mechanism.

Third-party papers, code and other excerpts in traces retain their own terms. The dataset-level notice does not relicense those materials, override their permissions, or resolve the outstanding third-party redistribution review. The privacy verification reports are not comprehensive security or legal clearance. release.json records publication metadata; historical verification/redaction reports retain their original scope and results.

Contents

1,019 original trace identities and 40 labelled derived variants, across 70 recorded run IDs and 2026-08-19 through 2026-09-22 UTC. These are not a single comparable cohort or 1,059 independent experiments. Model/date/protocol inventories and all recorded errors and cap hits are retained. Missing measurements remain missing.

The JSONL trace and linked call/node/tool/edit/telemetry/metric tables are retained, as are CSV indexes and a freshly rebuilt SQLite database. Raw objects, source logs, paper/reference archives and source-restoration tools are deliberately NOT included. Source/archive inventories are provenance metadata only, not downloadable payloads.

Identity and measurement semantics

record_id remains the ORIGINAL content hash for stable joins; it is no longer the hash of the redacted trace. sanitized_trace_sha256 verifies the redacted payload. Episode/source IDs and protocol/config/artifact hashes also refer to ORIGINAL data. Never execute sanitized commands or treat anonymized paths as runnable locators.

All token counts, timing, scores, LOC and other numerical measurements describe the original pre-redaction executions. They are preserved, not recalculated from the edited text. Redacted text will not necessarily retokenize to those counts. Canonical originals and derived variants remain distinct; infrastructure failures are not model-quality zeroes. Compare only compatible frozen protocol/task/config cohorts. Behavioral signatures do not identify undisclosed training coefficients.

Verification and reuse

Recorded validation results are in verification.json. This release contains data and documentation only; maintenance scripts are not included.

This is NOT public-release clearance: third-party licenses, copied paper/code text, benchmark exposure, other identifiers and applicable provider terms still need review. Pattern-based privacy cleaning cannot guarantee exhaustive anonymization. The source archive's credential-like alerts were not confirmed live credentials.