localgate/audit-results

LocalGate — Human Audits and Judge Calibration

Last updated: 2026-09-07. The three human evaluation stages supporting LocalGate's open-ended MMLU-Pro study: conversion decisions, rewrite fidelity, and judge calibration. Each stage used three study authors as annotators. These are sampled instrument checks, not human labels for every question in the benchmark.

The four main study datasets remain separate: closed multiple-choice scores, open-ended questions, open Gemma generations, and bulk judge verdicts. This repository brings their human audit evidence together without mixing audit samples into the benchmark's training or test tables.

Load the data

The repository is private. Authenticate with an account that has access; token=True uses your configured Hugging Face credential.

from datasets import load_dataset

repo = "localgate/audit-results"
conversion = load_dataset(repo, "pass1", split="train", token=True)
fidelity = load_dataset(repo, "pass2", split="train", token=True)
human_calibration = load_dataset(repo, "calibration_items", split="train", token=True)
judge_calibration = load_dataset(repo, "calibration", split="train", token=True)

All configurations use train as an archival storage label. It does not mean these items are an independent training partition. Pin a Hub commit SHA with revision= for reproducible analysis.

Three stages, four tables

Stage	Configuration	Rows	What one row represents
1. Conversion decision	pass1	248	One original question with three human verdicts on convertibility and reference usability
2. Rewrite fidelity	pass2	152	One retained rewrite with three human verdicts on fidelity and self-containedness
3. Judge calibration	calibration_items	100	One question-response pair with three human verdicts and derived consensus
3. Judge calibration	calibration	300	One of three LLM judges' evaluations of a calibration pair

The first two stages join on blind_id or question_id; all 152 pass-2 questions are retained members of the 248-question pass-1 draw. Calibration joins on key, which identifies a (question_id, rep) pair. The calibration sample contains 100 items, not 300 independent examples: the latter count is three judge evaluations per item.

Human raters use consistent ordinal fields human_1, human_2, and human_3. Annotator names and freeform human notes are omitted from the published tables. Recorded verdicts are preserved, as are unsure flags in the first two stages; uncertainty flags do not change the scoring rules. Calibration's 100 answer keys come from 90 distinct questions.

Results

The audits closed on September 1, 2026. Metrics were recomputed using the study's scorers from the final Space exports.

Stage	Main result	Interpretation
Conversion decision	Population-weighted disagreement 21.59%; Fleiss' κ 0.280	Compares human majority with the converter; 7 items had no yes/no majority
Rewrite fidelity	Population-weighted same-question error 4.46%; Fleiss' κ 0.572	Passed the registered ≤5% gate, with a narrow margin
Judge calibration	Panel Scott's π 0.79992	Below the registered 0.80 gate; bulk judging proceeded under a documented deviation

Conversion and fidelity estimates account for deliberately unequal stratum sampling. They are not raw proportions over the displayed rows. The calibration sample also overrepresents selected edge cases and is not a representative estimate of performance across all bulk responses.

Two limitations travel with these results. A single additional majority rejection in the largest fidelity stratum would raise its estimate to 5.41%. Separately, human rater 2 misread the polarity of the pass-2 self_contained question. Those grades are preserved as recorded and must not be treated as valid evidence about self-containedness. Details and sensitivity checks are in conversion and fidelity. The calibration gate deviation and full human-versus-judge analysis are in calibration.

Stage 1 and 2 fields

Both tables retain the original question, options, reference, and blind mapping so a reader can inspect an item without reconstructing it from another dataset.

Field	Meaning
question_id	Original MMLU-Pro integer identifier; join key to the core datasets.
blind_id	Audit identifier such as A002; it links the two passes and the published audit key.
stratum	Sampling group; pass 2 uses borderline_kept for retained items from the borderline stratum.
stratum_population	Number of corpus items represented by that stratum for the corresponding stage's weighting.
category	Source MMLU-Pro subject.
question, options, reference_answer	Original stem, ordered answer choices, and correct-option text.
machine_convertible	Converter's original final keep/drop decision; not a machine judgement of rewrite fidelity.
human_1_unsure … human_3_unsure	Recorded uncertainty flags; independent of the selected verdict.

Additional pass1 fields:

Field	Meaning
human_1_converts … human_3_converts	yes, no, or borderline: would the original question retain one reference answer without its options?
human_1_answer_stands_alone … human_3_answer_stands_alone	yes, no, or borderline: is the reference usable as a free-text grading target?
majority_converts	Boolean yes/no majority from at least two humans; null if neither side has two votes.
majority_answer_stands_alone	The same majority rule for reference usability.

Additional pass2 fields:

Field	Meaning
open_question	The actual retained rewrite shown in the second pass.
human_1_same_question … human_3_same_question	yes, no, or borderline: does the rewrite preserve what the original asked?
human_1_self_contained … human_3_self_contained	Whether the rewrite can be answered without options; retain the rater-2 polarity caveat.
majority_same_question, majority_self_contained	Boolean yes/no majority, or null if unresolved; the latter inherits the polarity problem.

The calibration field dictionary and exact rubric are in calibration.md. match is the positive binary verdict there; no_match and no_answer are both negative. Binary consensus and three-way consensus are different quantities.

Collection and provenance

Pass 1 was blinded to the converter's decision and rewrite. Annotators saw the source question, options, and correct answer in different orders. Pass 2 followed completion of pass 1 and revealed the rewrite. Human judgements did not silently replace machine conversion decisions in the open corpus.

The final source files are verdicts_*.json for pass 1 and conversion_*.json for pass 2, reduced from append-only Space event logs by the study's loader. The old verdicts2_*.json files belong to the superseded pre-Spaces workflow and are not results used here. Calibration uses the final annotate_*.json files and the three corresponding judge runs.

The loader's provenance checks have limits: one annotator export was checked against its local clone's identity rather than independently against Hub commits, and another lacked git history. The published values are the retained exports, not independently attested records of every annotation event.

Supporting files:

• Audit key: blind-ID mapping and converter stage, score, reason, failure tag, and reference-answer flag.
• Population strata and sample provenance.
• Export provenance and the exact original audit rubric.
• Pass-1 metrics and pass-2 metrics.
• Conversion/fidelity rubric and scoring; calibration details.

Use and limitations

Use this release to inspect the audit, reproduce its scoring, and study human-model agreement. Three author-annotators are a small, nonrepresentative panel. Stratified sample results, chance-corrected agreement, and gate decisions answer different questions; none certifies every benchmark item or establishes general judge accuracy. Machine explanations may be wrong, and original benchmark content may include errors, names, or sensitive topics. Freeform human notes are not available in these tables, limiting qualitative reconstruction.

The open corpus's frozen study partitions remain unchanged. Audit membership does not create a separate leakage-free training set; source exclusions leave three identical-rewrite groups crossing active partitions, as documented in the open dataset card.

Licence and citation

This derivative release retains MIT metadata, matching the MMLU-Pro source card. Consult source licences and model terms for their respective materials. Cite MMLU-Pro and identify the LocalGate dataset URL and revision used.

@misc{localgate2026audit,
  author = {{LocalGate}},
  title = {LocalGate Human Audits and Judge Calibration},
  year = {2026},
  url = {https://huggingface.co/datasets/localgate/audit-results}
}

For corrections, open a dataset discussion with the configuration, item identifier, and revision, or use the LocalGate issue tracker.