AmoghSingh123/flipeval-artifacts

FlipEval artifacts — v1.0.0

Companion release for Certifying Compressed Language Models: An Audit and a Statistical Toolkit.

The paper's headline audit finding is that 0 of 17 published near-lossless claims release the per-item outputs a third party would need to run the paired comparison themselves. This repository is that recommendation applied to its own author: every per-item output behind every controlled result in the paper is here, in full, alongside the integrity metadata needed to check it.

Canonical citation is the Zenodo DOI (see Citation below). This repository is a convenience mirror of the data; the DOI is what the paper cites.

What is here

Directory	Contents
per_item_outputs/	88 cell JSONLs, extracted and directly loadable. One row per evaluation item, per model × method × seed × task.
archives/	The two sealed tarballs, their .sha256, and per-file manifests. Byte-identical to what was archived at run time.
atlas/	Rev-2 per-cell paired statistics, the exclusion table, the frozen 59-pair manifest, and the identical-score churn table. Derived statistics only — see Licensing boundary.
audit/	The frozen 17-claim table (with recorded source content hashes) and the per-claim verdict CSV with every robustness and transparency column, including R04's set-aside GSM8K computation.
certification/	Required-$n$ certification tables at all three margins (1, 2, 3 pp).
h3/	The nine eight-cell H3 artifacts, both grid validation summaries, and the escalation screen summary.
reproduction/	Preregistration, all registrations with dated amendments, signed decision records, configs, container image sha256 + build recipe, SLURM scripts, build receipts, freeze fingerprints, and the campaign incident log.

Both forms of the cell data, deliberately

per_item_outputs/ is extracted so the data is previewable in the dataset viewer and loadable with datasets. archives/ carries the same bytes sealed, with per-file SHA-256 manifests, so the extracted copy can be checked against the record made at run time. Usable and integrity-checkable is the pairing the paper argues for; shipping only one of the two would concede half the point.

Paths under per_item_outputs/ retain the results/<run_dir>/ prefix recorded in the archive manifests, so manifest paths resolve directly against the extracted copy.

# verify the archives, then verify the extracted copy against them
sha256sum -c archives/*.sha256
python - <<'PY'
import json, hashlib, os
for a in ["minigrid_run_20260722", "escalation_run_20260726"]:
    man = json.load(open(f"archives/{a}.manifest.json"))["files"]
    items = man.items() if isinstance(man, dict) else [(f["path"], f) for f in man]
    bad = [p for p, r in items
           if hashlib.sha256(open(f"per_item_outputs/{p}", "rb").read()).hexdigest()
              != (r["sha256"] if isinstance(r, dict) else r)]
    print(a, "OK" if not bad else f"MISMATCH: {bad}")
PY

Licensing boundary — read before redistributing

• Everything in this repository is released CC-BY-4.0 and is ours to release: the per-item outputs were generated by us, and the atlas files are derived per-cell statistics we computed.
• Raw upstream per-item files from stratum S1 are not redistributed here, and must not be added. S1 is built on the Open LLM Leaderboard v1 archive, which carries no declared license. This is recorded as a limitation in the paper and in the datasheet, not worked around.
• A third party can re-derive S1 from the original sources. The frozen manifest atlas/atlas_pair_manifest.json records, per pair, the upstream dataset identifiers and run timestamps needed to fetch the same inputs and recompute every derived column with the released flipeval package.
• Stratum S2 (vendor-published evaluations) is Apache-2.0 upstream.

Redaction

These runs executed on a university HPC cluster. Site-specific identifiers — absolute paths, the SLURM charge account, usernames, login and compute hostnames — were replaced with placeholders before publication. reproduction/REDACTIONS.json states the rules and which files were touched.

No scientific quantity was altered. No accuracy, hash, count, seed, timestamp, job ID, model revision or decision value is affected, and the sealed archives were not rewritten at all — which is why their recorded SHA-256 values still verify. Compute nodes are pseudonymised one-to-one and stably, so incident 11's finding (the node SLURM was asked to exclude differed from the node the job landed on) remains checkable from the pseudonyms alone.

Provenance and integrity

• Container image SHA-256: 8260d04cf1f76cb5961b6538dbeb9178006b29c5d8c93d2c639976bcd1db2007 (reproduction/image/ carries the build recipe; the image itself is not redistributed).
• Source-state fingerprint: reproduction/fingerprints/PREPACE_FREEZE.json.
• reproduction/CAMPAIGN_INCIDENT_LOG.md documents 28 defects caught during the campaign, with what each one would have corrupted. It is included on purpose. A pipeline with 28 recorded catches is better evidence of verification than one presenting a clean surface, and several entries — a validator pointed at the wrong grid, a write-protection scheme that could not stop os.replace(), an sbatch flag accepted and silently discarded — are reusable warnings for anyone running a comparable campaign. Entries 27 and 28 are about this release: 27, that the first scrub pass recorded its own reverse mapping, which would have de-anonymised every identifier it replaced; and 28, that the source repository was public for eight days while every project document asserted it was private, because nobody ran the one command that would have checked.

Limitations

• The atlas is a record, not a sample: it describes the evidence the field circulates, not the behaviour of quantization in general.
• All benchmarks are English and predominantly multiple-choice or short-answer. Float-scored generation tasks are outside the flip model entirely.
• The controlled experiment is 8 cells, 1.5B–8B, 4-bit, two methods, five seeds. The GSM8K arm at n = 1,000 is under-resolved, which the paper states and quantifies rather than omits.
• No human-subject data and no personally identifying information.

Revision history

Both revisions are published, and the delta is reported. Rev-1 contained a population defect: an omitted reverse-chronological fallback clause and a parser that missed a newer nested metrics schema had silently dropped cells, the first non-randomly. An independent spot-check reconciled all 262 compared fields — the arithmetic was right and the population was not.

Rev-2 files are the ones the paper cites and are what this release carries as current. Rev-1 files remain in the source repository and in the git history at the tagged commit. A corrected artifact whose correction is visible is the paper's own argument applied to itself, so the superseded revision is retained rather than overwritten.

Citation

Cite the archived release (canonical):

@dataset{singh2026flipeval_artifacts,
  author    = {Singh, Amogh},
  title     = {{FlipEval} artifacts: per-item outputs, flip atlas, audit and
               certification tables},
  year      = {2026},
  version   = {v1.0.0},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.21708923}
}

The DOI above is the v1.0.0 version DOI, which is what the paper cites: it resolves to this frozen state. The concept DOI 10.5281/zenodo.21708922 resolves to whichever version is latest, and is the right one to cite when you mean the artifact series rather than this release.

Code: https://github.com/amoghsingh130/flipeval (Apache-2.0), tag v1.0.0.