localgate/judge-panel-verdicts

MMLU-Pro Open: judge panel verdicts

Three LLM judges independently graded each of 42,270 Gemma 4 E2B responses (8,454 questions × five attempts). This release contains 126,810 verdicts, and their per-question majority scores.

These are reference-based correctness assessments, not human ground truth.

Place in the study

Closed multiple-choice baseline → open-ended questions → Gemma generations and performance → judge verdicts and consensus.

Join questions using question_id. Join individual bulk answers using question_id plus rep here, equivalent to attempt in the generations dataset. key is the string question_id:rep.

Load the data

Last updated: 2026-09-07.

from datasets import load_dataset

repo = "localgate/judge-panel-verdicts"
labels = load_dataset(repo, "consensus_labels", split="train", token=True)
judgements = load_dataset(repo, "judgements", split="train", token=True)

For a private repository, authenticate with an account that has access. Pin a Hub commit with revision= when reproducing results.

Configuration	Rows	Unit
consensus_labels	8,454	One question, aggregated over five attempts
judgements	126,810	One judge's verdict on one bulk answer

Here train is an archival storage partition, not the study's training set. The frozen question membership in provenance/split.json separates 5,851 training, 857 validation, and 1,655 test questions; a further 91 are excluded from split-based analysis by the stem-leak repair. All 8,454 questions remain available for archival and whole-corpus analyses.

How the labels were made

The panel used AWS Bedrock model identifiers mistral.ministral-3-14b-instruct, openai.gpt-oss-120b-1:0, and qwen.qwen3-next-80b-a3b. Each received the question, authoritative reference answer, and Gemma response, without other judges' verdicts. Settings were temperature 0, a 400-token judge-output cap, and prompted JSON; GPT-OSS used reasoning_effort=low. See the full protocol for exact prompt text, schema, fingerprints, and retry/escape handling.

match maps to 1; no_match and no_answer map to 0. At least two of three positive votes make an attempt correct. correct_count sums those five attempt outcomes; k is always 5, the number of attempts, not the score.

Bulk measurement	Result
Complete three-judge panels	42,270 / 42,270
Fleiss' κ, binary correctness	0.8309
Fleiss' κ, three-way verdict	0.7605
Exact three-way unanimity	33,473 / 42,270 (79.19%)
Individual verdicts: match / no_match / no_answer	59,465 / 53,405 / 13,940

Unanimity means all three judges chose the same three-way verdict. These agreement measures describe consistency among judges, not their accuracy.

Calibration: panel Scott's π was 0.79992, below the planned 0.80 gate; bulk judging proceeded under the documented deviation in the separate audit release.

Data fields

judgements

Field	Meaning
key	Answer key, question_id:rep; unique only together with judge
question_id	Integer join key inherited from MMLU-Pro
rep	Gemma attempt index, 0–4
judge	Full bedrock: provider/model specification
judge_model	Short label: ministral, gptoss, or qwen
verdict	match, no_match, or no_answer under the published rubric
verdict_binary	1 for match; 0 otherwise
reason	Judge-generated explanation, not a verified rationale
tokens_in, tokens_out	Recorded input/output token usage for the successful grading operation
finish_reason	The Gemma generation's finish signal, not the judge's
prompts	System-prompt-plus-schema digest, e2c597632659
template	User-message-template digest, ce0113675bf4
output_mode	prompted JSON
reasoning_effort	low for GPT-OSS; null means unspecified for the others

consensus_labels

Field	Meaning
question_id, category	Question identity and source subject
correct_count	Number of attempts judged correct by binary majority, 0–5
no_answer_count	Number of attempts where any judge said no_answer, 0–5
k	Number of attempts, always 5
list_style	Whether the stripped original stem starts with “which of the following”, ignoring case

no_answer_count can overlap with correct_count: one judge may say no_answer while the other two say match. It is not a count of majority nonanswers, refusals, or token-limit truncations.

Compute, uses, and limits

The retained bulk records contain 180,130,500 input tokens and 8,142,073 output tokens (188,272,573 total). These are recorded grading-operation usage totals; they exclude separate calibration runs and discarded failed operations. They are not a complete provider invoice or a measured energy/carbon footprint.

Use this release to reproduce the panel scores, inspect disagreements, or compare reference-based graders. The upstream questions were selected by conversion filters, the source reference can itself be flawed, and judges can share errors. Model explanations and no_answer verdicts require interpretation, even when the provider reports a normal generation stop. These results do not establish general grader reliability outside this benchmark. Keep question families and study split exclusions together when training or evaluating models.

Sources and citation

Maintained by the LocalGate project. This derived release retains the card's MIT license declaration; source datasets and model artifacts have their own terms. See MMLU-Pro, Ministral 3 14B, GPT-OSS 120B, and Qwen3-Next 80B A3B.

When using these data, cite the upstream MMLU-Pro paper and identify this dataset's Hub revision. A dataset citation is:

@misc{localgate_judge_panel_2026,
  author = {{LocalGate}},
  title = {MMLU-Pro Open: Judge Panel Verdicts},
  year = {2026},
  howpublished = {Hugging Face dataset},
  url = {https://huggingface.co/datasets/localgate/judge-panel-verdicts}
}