SoroushVahidi/mwfas-benchmark-extension

MWFAS Benchmark Extension: metadata and objective witnesses

Research author: Soroush Vahidi. This is a separate extension of the original MWFAS Heuristic Metrics dataset, prepared for an unpublished JDSA benchmark manuscript. It does not change or replace the original 7,371-row, 14-configuration release. No journal DOI or archival DOI is claimed.

Contents and observation units

• Five exact, path-free EXP18 projection tables under metadata/: 288 source-parent records, 442 graph rows under named normalization contracts, 5,254 run-outcome records, 243 numerical reference records, and 2,607 artifact-availability records. metadata/schema.json defines the fields and joins. These are metadata and recorded results; most underlying graph and order bytes are not included.
• Three exact historical Wikimedia-derived weighted directed graph bytes and ten complete saved orderings under data/wikimedia/. graph_inventory.csv and witnesses.csv pin their byte hashes, normalized graph/order hashes, vertex sets, parent, objective convention, statuses, and frozen expected backward-arc weights. These 13 files are one Wikimedia clickstream source parent with three graph views, not three independent datasets.
• CHECKSUMS.csv authenticates all published package files except itself. The independent GitHub verifier recomputes graph/order hashes, all ten backward weights, retained-DAG acyclicity, and metadata joins/status/bound invariants without optimization or proprietary solvers.

graphs.csv records 442 contract-scoped graph rows, not 442 globally distinct graphs. The common-contract core has 416 weighted graph identities across 317 topology identities. Weight realizations, views, seeds and repeated runs are not independent source datasets. EXP07's 45 references reuse nine EXP03 and 36 EXP04 graph bytes and add no graph or outcome row. The original public-dataset audit remains 50 PASS, 3 FAIL and 22 NOT_VERIFIABLE; this extension does not rewrite its rows or statuses.

Methods, provenance and exact objective operation

The author-generated synthetic cohorts were produced by pinned EXP03/04/06 designs and executed on NJIT Wulver; their raw graph and ordering bytes are not distributed here pending a specific institutional facility-use/rights determination. EXP05 includes external EXIOBASE and SNAP sources. EXP12 has 20 Ali graph references with no established collection-wide raw redistribution grant. EXP13 covers OpenFlights, Census LODES and IRS SOI under source-specific conditions. These source parents and access classes remain explicit in source_parents.csv and artifact_availability.csv; a metadata row is not an access promise.

The Wikimedia graph bytes are pinned derived artifacts at workspace commit 975b287f, copied byte-for-byte here after public SHA-256 recovery. Their conversion retained link rows of weight at least five and omitted self-loops. The vertex set is the union of graph arc endpoints; every saved ordering lists each vertex exactly once. Arcs point from source to target. For an ascending saved rank, a backward arc has rank(source) > rank(target), and its integer weight contributes to the active objective. The verifier also checks that the retained forward arcs form a DAG. The ten method labels are contextual historical labels; execution clocks for these saved rankings are not independently verified. The original pre-May-2026 Wikimedia monthly TSV snapshots and compressed-file hashes remain unavailable, and current publisher TSVs differ from these pinned derived graphs. Rescoring a pinned graph is not reconstruction from its original TSV or a new algorithm run.

EXP13 metadata use an active graph after self-loop removal. To compare against the original source objective, add 18,628 for LODES or 124,358,925 for IRS; OpenFlights has zero loop constant. Incomplete statuses have no assigned objective. The 198 EXP06 and 20 EXP07 NUMERICAL_SOLVER_CERTIFIED records state a recorded numerical solver certificate under their settings; the other 25 EXP07 FEASIBLE_WITH_BOUND cases retain lower/upper bounds and no optimum. None is an exact-arithmetic proof. Raw Gurobi logs, solver binaries and private storage paths are absent.

Reproduce from this public release

Use Python 3.12 or newer and the matching immutable dataset revision. Download it with hf download SoroushVahidi/mwfas-benchmark-extension --repo-type dataset --revision <revision> --local-dir extension. Clone the GitHub verifier release and run python release/verify_objective_witnesses.py extension. The verifier needs only standard-library Python; it does not rerun optimization. It checks every published graph and ordering byte and reconciles their ten objectives with the frozen metadata. It checks metadata arithmetic, statuses, joins, loop constants and numerical bound relationships, but cannot independently recompute objectives for private or reference-only graph/order artifacts.

The original Hugging Face dataset lacks its historical graph/order inputs. The later manuscript's central synthetic, Ali, EXP05 and EXP13 objectives therefore remain partly dependent on withheld, externally referenced or source-restricted witnesses. This public subset supports complete saved-byte reproduction of the ten Wikimedia objectives and metadata-level checks elsewhere; it does not yet close the paper's full central objective-witness gap.

License, attribution and citation

CC BY 4.0 applies to Soroush Vahidi's author-controlled metadata, original compilation, transformation and saved ranking contributions in this package; see exact scope and full terms. Wikimedia's Analytics Clickstream source is CC0; credit Wikimedia Analytics and Ellery Wulczyn and Dario Taraborelli, Wikipedia Clickstream (2015), DOI 10.6084/m9.figshare.1305770, as well as Vahidi's conversion, the language/month and this exact revision. External Ali, SNAP, EXIOBASE, OpenFlights, Census and IRS bytes are not relicensed or mirrored. A URL or hash in the metadata grants no redistribution rights to that external file.

Cite Soroush Vahidi, MWFAS Benchmark Extension: metadata and objective witnesses, this immutable dataset revision, and the original historical dataset when used. Cite the eventual JDSA manuscript only after publication. The prior ranking article is separate work.