LeoJiangOR/FrontierOR-Audited-180

license: cc-by-4.0 language:

• —en prettyname: FrontierOR Audited 180 sizecategories:
• —1K<n<10K task_categories:
• —other tags:
• —operations-research
• —optimization
• —mixed-integer-programming
• —gurobi
• —benchmark
• —llm-for-or
• —audited configs:
• —configname: meta datafiles: papermetainfo.json ---

FrontierOR Audited 180

This is an evaluator-oriented derivative of `SmartOR/FrontierOR`, pinned to upstream commit 37ccd8b6dca3bf7f4e0c58941a6ed156832a6d9e. The released descriptions, formulations, Gurobi implementations, solution schemas, reference solutions, and feasibility checkers were audited and repaired as one evaluation contract.

Final status

The compatibility alias is ostrowski2012 -> earl2005. The upstream ostrowski2012 directory was a mislabeled byte-level duplicate of the Earl/D'Andrea 2005 defensive-drill task and had no canonical metadata row. It is retained so that existing 180-directory evaluation code does not break, and its alias status is disclosed in the prompt, formulation, model, checker, and CASE_IDENTITY.json.

What “audited” means

For every published instance, its bundled solution is finite, accepted by the current checker, and has a bundled checker result matching a fresh replay. Objective-tamper preflight and decision-erasure mutation probes were also run. The nine checkers that formerly reconstructed omitted decisions with Gurobi now receive complete operational witnesses. They validate bounds, domains, constraints, and objectives by direct numeric substitution. Their 54-pair regression and mutation results are recorded in audit/SOLVERLESS_CHECKER_AUDIT.json.

The static triad scan is deliberately heuristic and still reports lexical findings; those findings are retained for transparency and are not runtime failures.

Optimality boundary

“Accepted” means a complete checker-verifiable feasible evaluation baseline. It does not mean that every historical large incumbent was re-proved globally optimal. For the 1,095 published references, the provenance classes are explicit_incumbent=398, solver_optimal=474, solver_optimal_with_tolerance=77, objective_without_certificate=138, ambiguous_status=8. In total, 544 published references across 128 cases do not carry a global-optimality certificate. No feasible incumbent has been relabeled as a proven optimum.

Layout

Each case directory contains its description, formulation, schemas, generation code, checker, official instances, paired solutions, paired feasibility results, and any matching solver logs. Cases using generic solver-free witness checking also contain trusted compressed coefficients under solverless_models/. Gurobi can still be used offline to regenerate reference solutions, but it is not needed to run any released feasibility checker. Two legacy segundo2019 solutions without corresponding instances are excluded from this release.

• —RELEASE_MANIFEST_180.json: case identities and all 1,095 official pair bindings.
• —audit/: final per-case matrix, change manifest, and compact gate summaries.
• —PUBLISH_FILE_MANIFEST.json and SHA256SUMS: complete file inventory and hashes.

Metadata can be loaded with:

from datasets import load_dataset

meta = load_dataset("LeoJiangOR/FrontierOR-Audited-180", "meta", split="train")
print(meta[0])

Download the complete self-contained task tree with:

hf download LeoJiangOR/FrontierOR-Audited-180 --repo-type dataset --local-dir frontieror-audited-180

Provenance and license

• —Upstream dataset commit: 37ccd8b6dca3bf7f4e0c58941a6ed156832a6d9e
• —Evaluator commit: 93fb106d98c9d21729aa846f29822e12ad22db90
• —Audit/release date: 2026-08-29 (America/New_York)

The upstream dataset card declares CC-BY-4.0. This derivative retains that license and attribution; see LICENSE.md and UPSTREAM_README.md.