OPERATE

Benchmarking Persistent Operational Agency in Source-Grounded Executable Systems

OPERATE evaluates LLM agents as long-running operational decision centers in seeded executable systems. The environment, not the model, produces state transitions. Observations are partial, events and actions are typed, and scores are linked to recorded evidence.

This public dataset is the runtime companion to the single current state of the OPERATE code repository. It is public, ungated, and intentionally has no selectable public version series. It belongs to the OPERATE collection. For a formal run, record the exact 40-character HF commit SHA shown by the Hub.

Updates accumulate as commits; the current tree contains only the latest Full, Lite and runtime artifacts. See the repository update log.

Benchmark scope

The Core contains 769 scenarios over 502 physical sources and seven domains. OPERATE-Lite contains 193 exact Core-locked rows from 122 physical sources. It retains a 104-row coverage core over joint task classes, source families, event/control mechanisms, native scale and declared source variation. Complete rounds then increase independent-source support, adding 23, 14 and 13 rows until the first complete round inside the 150–200-row development budget. All admitted Autonomous Driving, Building Energy, Microgrid, Power Grid and Traffic rows are then retained, adding 30 window/condition variants. Datacenter retains all 11 medium and 7 high cases plus its 9 selected basic cases, adding 9 rows; Logistics remains at 66 selected rows. Every row has an inclusion/exclusion reason. Selection does not use LLM scores. Core admission supplies the quality requirement, not the size budget. All 17 backends, 22 task families, four difficulty levels and six horizon buckets remain covered. This is a development/ablation subset, not a statistical sample, a mathematical minimum or the Full/Core leaderboard denominator.

Primary results use logical_persistent. realtime_persistent is a separate supervision treatment for proactive monitoring, correct silence, latency, cancellation, supersession, action lifecycle, and safety takeover. logical_stateless is a compatibility ablation.

Files

File	Purpose
MANIFEST.json	Exact file hashes and required runtime/source bindings
backend_runtime_closure.json	Runtime packages, archives, links, and external sources
full/test-00000-of-00001.parquet	Self-contained Full scenario contracts and suite metadata
lite/test-00000-of-00001.parquet	Self-contained Lite scenario contracts and suite metadata
parquet_manifest.json	Full/Lite Parquet hashes, row counts, and catalog bindings
backends.tar.zst	Redistributable native runtime assets

Scenario contracts, Full/Lite definitions, source locks, evaluation code, and install tooling live in GitHub. This companion restores the large redistributed assets omitted from Git. Assets distributed through their upstream repositories remain URL- and checksum-bound in the manifests.

The Parquet configurations are independently browsable and reversible. Each row contains the exact scenario YAML and row metadata; shared suite metadata is stored once in the original Parquet file, not repeated in the viewer. See the Parquet schema. The 22 public columns omit redundant release IDs, track labels, and admission status fields. Full/Lite is identified by subset. Internal reference IDs remain inside the reversible JSON/YAML payloads and runtime manifests where required for exact reconstruction and cross-file checks; they are not selectable public tags.

Historical qualification does not have to match the code of a new independent evaluation. Each run records its actual implementation; data integrity, same-run stability and compatible resume/merge remain mandatory. Maintenance uses affected-scope tests rather than automatically repeating full calibration.

from datasets import load_dataset

full = load_dataset("Xnhyacinth/OPERATE", "full", split="test")
lite = load_dataset("Xnhyacinth/OPERATE", "lite", split="test")

Download and verify

Download the current public snapshot:

git clone https://github.com/Xnhyacinth/OPERATE.git
cd OPERATE
python -m pip install uv==0.12.5
uv sync --frozen --python 3.13 --extra dev --extra llm --extra hf \
  --extra released-backends --extra simulators

uv run python scripts/download_from_hf.py --download-only

Anonymous download is supported; no HF_TOKEN is required. The local operate_data/ directory is not a selectable benchmark version; MANIFEST.json binds the installed bytes to the current Core, and the local owner receipt records the resolved immutable HF commit. Pass --revision <HF-COMMIT-SHA> for a pinned reproduction.

The runtime bundle includes only the three byte-exact M5 tables referenced by Core. Their hashes and 81 scenario bindings are recorded in MANIFEST.json. The original source_lock.json is carried separately as metadata, without adding a fourth physical input. NGSIM bundles also retain their original checksums.sha256 lists for native verification. On 2026-09-04, the dataset publisher confirmed that it holds permission to redistribute these files. The M5 Competition Rules remain applicable; OR-Gym code remains MIT-licensed. No M5_ZIP or KAGGLE_TOKEN is required for the bundled snapshot.

The immutable release closure retains its admission-time external-acquisition record. The top-level source_assets.m5 entry in MANIFEST.json is the authoritative distribution-time overlay for the now-bundled files.

For complete native installation and a baseline runtime check, run:

bash scripts/setup_eval_env.sh --smoke

The setup script acquires the remaining declared sources and runs one wait_only episode per released backend. Required-source and smoke failures are fatal. This validates runtime/evidence integrity, not model performance; model evaluation additionally requires your model API credentials. Unanswered passive-baseline alarms remain reported as policy warnings, not successful interventions; default strict evaluation is unchanged.

See the GitHub README for a baseline episode and Lite command, and the formal evaluation runbook for treatment-bound runs.

Intended use and limitations

OPERATE is intended for evaluating persistent operational agency, tool use, monitoring, and intervention under partial observability. It is not a training corpus, a safety certification, or evidence that a model is suitable for live infrastructure control.

• Domain row counts are uneven. Official aggregation is stratified; raw row counts do not measure physical-source diversity.
• Lite substantially undersamples Logistics sources and must be reported under its own track name.
• Procedural stressors are seeded and labelled; they do not become real source events merely because the base state is source-grounded.
• Public reproducibility and leaderboard eligibility are separate. Formal provider evaluations and result distribution are still pending, so official leaderboard submissions are not open.

Provenance, licenses, and citation

OPERATE-authored code and metadata are MIT-licensed. This companion is a mixed-license collection: every redistributed upstream asset retains its own terms, notices, and per-root manifest binding. The other metadata value does not relicense those assets under MIT; this avoids relicensing upstream data under MIT. See the third-party license inventory and data provenance.

When reporting results, cite the repository plus the exact Git commit, HF revision, treatment hash, and model/provider binding. Citation metadata is provided in CITATION.cff.